A self-adaptive day lily segmentation and four-dimensional picking point positioning method based on heterogeneous dual-domain cooperation

Abstract

The present application aims at the problems of dense occlusion, high similarity of neck leaves, high dependence of artificial design of semantic segmentation model, and low positioning accuracy of picking points in day lily picking scene, constructs an automatic architecture method of H²D²-MONAS segmentation model, integrates heterogeneous stage search, high degree of freedom topology coding, dual domain adaptive fusion congestion calculation, cell level cooperative evolution operator, multi-dimensional cooperative evaluation and nonlinear geometric fusion mechanism, and searches for a high-precision lightweight segmentation model suitable for the scene; then the MW-GCA algorithm is used to obtain the two-dimensional pixel coordinates and the cutting angle of the picking point, and the four-dimensional coordinate conversion and the kinematics analysis algorithm of the mechanical arm are used to generate the executable control instructions of the mechanical arm, so as to realize the automatic and accurate picking of day lily. The present application effectively improves the segmentation, positioning and attitude control precision, reduces the plant damage rate, and balances the model lightweight and search efficiency, and provides a complete technical solution for the engineering landing of intelligent day lily picking.

Description

Technical Field

[0001] This invention belongs to the field of intelligent crop harvesting technology, specifically an adaptive daylily segmentation and four-dimensional harvesting point positioning method based on heterogeneous dual-domain collaboration. Background Technology

[0002] With the rapid development of deep learning technology, semantic segmentation technology is increasingly widely used in the field of intelligent agricultural equipment. Among these applications, the automated design of semantic segmentation models for daylilies has become a key support for promoting the transformation and upgrading of the daylily industry. Daylilies are perennial herbaceous plants of the Narcissus family, belonging to a distinctive economic crop with both medicinal and edible uses. my country is the world's largest producer, ranking first globally in both planting area and annual output. However, the current daylily harvesting process faces significant technical bottlenecks: harvesting is highly dependent on manual labor and is subject to strong seasonality and time sensitivity. The harvesting period is concentrated in the hot and rainy months of June to August, resulting in significant spatial and temporal constraints, labor shortages, and delayed harvesting leading to quality degradation and severe economic losses. This predicament has driven the research and development of intelligent daylily harvesting technology, with automated model construction serving as its core support, requiring both low latency and high robustness. Neural Architecture Search (NAS) can autonomously generate the optimal network architecture adapted to the semantic segmentation task of daylilies, providing technical support for the engineering implementation of intelligent harvesting. To address the challenges of dense occlusion, high similarity of neck leaves, and difficulty in classifying daylily harvesting scenes, traditional manually designed semantic segmentation networks rely on human experience, making them ill-suited to task requirements and unable to balance feature extraction effectiveness with real-time model performance, resulting in insufficient accuracy in practical applications. In contrast, neural network models generated by Neural Networks (NAS) leverage multi-level feature abstraction from convolutional neural networks and NAS's self-optimized feature fusion strategy to achieve pixel-level to scene-level semantic collaborative parsing. This results in superior robustness and accuracy in complex scenes with high-density clusters and multiple visual interferences, balancing model complexity and performance. However, current NAS-based semantic segmentation for daylilies still faces three core challenges: how to design a search space adapted to daylily harvesting scenarios, how to formulate efficient search strategies to balance performance and cost, and how to construct efficient evaluation strategies to improve NAS convergence speed. These challenges directly determine the adaptability and engineering feasibility of NAS technology in intelligent daylily harvesting. At the search space design level, existing general-purpose NAS search spaces fail to address the core pain points of daylily, such as dense occlusion, high neck-leaf similarity, and difficulty in class differentiation. They cannot meet the requirements for occlusion feature completion and fine-grained neck-leaf differentiation. A customized search space with fine-grained feature capture and edge enhancement for the daylily scenario needs to be constructed, balancing specificity and simplicity. At the search strategy design level, for a customized search space for daylilies, traditional NAS suffers from low search efficiency and susceptibility to local optima, making it difficult to balance segmentation accuracy, model lightweighting, and real-time field operations. An efficient search method adapted to the daylily harvesting scenario needs to be designed to quickly discover the optimal architecture. At the candidate architecture evaluation strategy design level, traditional evaluation has high computational overhead and slow convergence. Existing lightweight methods are not suitable for the complex environment of daylily fields. An efficient evaluation strategy needs to be designed based on the characteristics of the daylily dataset, balancing evaluation accuracy and low overhead, and improving the convergence speed of NAS. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide an adaptive daylily segmentation and four-dimensional picking point localization method based on heterogeneous dual-domain collaboration, which is mainly used for automated picking of daylilies in real-world scenarios. To achieve the above objectives, the present invention provides the following technical solution: An adaptive daylily segmentation and four-dimensional picking point localization method based on heterogeneous dual-domain collaboration is proposed. This method constructs an automatic architecture method for the H²D²-MONAS daylily segmentation model. The H²D²-MONAS model includes a hierarchical heterogeneous multi-source fusion search space for daylily segmentation, a multi-objective neural architecture search strategy guided by dual-domain semantic collaboration, and a low-cost neural architecture performance evaluation strategy based on multi-dimensional collaboration and ranking-driven approaches. The optimal daylily segmentation network architecture, obtained through joint optimization of the search space, search strategy, and evaluation strategy, is at the core. The MW-GCA algorithm is used to accurately acquire key information for daylily picking, and a robotic arm execution system completes the automated picking operation of daylilies. The hierarchical heterogeneous multi-source fusion search space for daylily segmentation includes an encoder-decoder end-link heterogeneous phased search unit for daylily segmentation tasks, and a high-degree-of-freedom topological coding strategy supporting single-node multi-precursor node connections; the dual-domain semantic collaborative guided multi-objective neural architecture search strategy includes a best point set initialization method, an adaptive fusion congestion calculation mechanism between the target performance domain and the architecture structure domain, a dual-mode crossover operator adapted to cell-level segmented coding, and a topology-operation collaborative mutation operator; the low-cost neural architecture performance evaluation strategy based on multi-dimensional collaboration and ranking includes four core evaluation indicators: model perturbation robustness, feature space isotropy, parameter-gradient joint uncertainty, and gradient-layer structure regularity, as well as a ranking-based nonlinear geometric fusion mechanism; the MW-GCA includes a Mini-Window local window mechanism, a VASU-based "U"-shaped feature window filtering algorithm, and a pruning angle calculation module; the robotic arm execution system includes a robotic arm kinematics analysis module and a four-dimensional coordinate transformation algorithm module. A further optimized method for constructing an automatic architecture for the H²D²-MONAS daylily segmentation model is as follows: First, addressing the core segmentation challenges of dense occlusion in daylily harvesting scenarios leading to feature loss, high visual similarity between flower stems and leaves, and difficulty in accurately distinguishing categories, this invention innovatively designs a hierarchical heterogeneous cell structure and a high-degree-of-freedom topological coding strategy adapted to this scenario. Through a phased Normal / Attention adaptation unit, features of occluded regions are preserved while suppressing background interference. A Feature Pyramid Aggregation (FPA) decoding unit enables multi-scale feature hierarchical aggregation, supplementing missing occluded information and enhancing the distinguishability of neck and leaf features. Simultaneously, relying on a single-node multi-preorder connection mechanism, residual skip connections, multi-branch parallel processing, and cross-level feature fusion are achieved, accurately extracting the subtle textures of daylilies and collaboratively optimizing segmentation accuracy and operational efficiency. Secondly, considering the significant differences in the growth states of daylilies in the field and the practical need for lightweight intelligent harvesting equipment to balance segmentation accuracy and model lightweighting, traditional multi-objective NAS is prone to damaging the customized cell structure of the scene, resulting in homogeneous solution sets and low optimization efficiency. This invention proposes a multi-objective neural architecture search strategy guided by dual-domain semantic collaboration: it improves population diversity by initializing the optimal point set to adapt to the recognition of daylilies in different growth states; it constructs an adaptive fusion crowding degree calculation mechanism between the target performance domain and the architecture structure domain to balance the accuracy of flower recognition, neck and leaf differentiation, and the adaptability of customized architecture; and it is combined with a cell-level co-evolutionary operator to improve search efficiency while preserving the integrity of the core structure, quickly selecting a high-precision lightweight segmentation architecture suitable for the field scene. Finally, addressing the pain points of daylily field applications such as large fluctuations in light intensity, uneven plant growth, and limited computing power of harvesting equipment, existing training-free NAS evaluation suffers from problems such as single-dimensionality, poor prediction reliability, and slow convergence. This invention proposes a low-cost neural architecture performance evaluation strategy driven by multi-dimensional collaboration and ranking. It comprehensively evaluates four dimensions: perturbation robustness, feature isotropy, generalization ability, and structural regularity, fully adapting to the complex field environment and lightweight deployment requirements. It introduces a ranking-based nonlinear geometric fusion mechanism to accurately predict model performance with low computational overhead, significantly shortening the NAS iteration cycle and efficiently discovering high-quality segmentation architectures suitable for intelligent daylily harvesting. After completing the construction of the three core NAS modules, this invention uses the NAS framework to optimize the optimal semantic segmentation model adapted to the daylily harvesting scenario, achieving high-precision semantic segmentation of daylily harvesting scenario images. In addition, based on the segmentation results, the MW-GCA intelligent localization algorithm for daylily harvesting points is further designed. By creating a "U"-shaped local window to filter corner information, and based on this window, an attitude line localization method is proposed, ultimately achieving accurate estimation of the pixel coordinates and angles of the harvesting point. Finally, this four-dimensional harvesting information is transformed into three-dimensional spatial coordinates and combined with the kinematic analysis algorithm of the robotic arm to complete the automated and precise harvesting of daylily plants.

[0004] The technical solution provided by this invention has the following advantages compared with the prior art: First, addressing the issues of feature loss due to dense occlusion and difficulty in fusing multi-scale features of the neck and leaves in daylily harvesting scenarios, and the inability of existing general-purpose NAS search spaces to adapt to this scenario and provide suitable feature support, this invention designs a phased Normal / Attention adaptation unit and a Feature Pyramid Aggregation (FPA) decoding unit for multi-scale fusion within the search space. The former achieves feature preservation and interference suppression under dense occlusion, while the latter, through multi-scale feature hierarchical aggregation, completes the missing features in the occluded region and enhances the distinguishability of features at different scales of the neck and leaves, further improving the accuracy of semantic segmentation of daylilies and solving the segmentation problem of high neck and leaf similarity and dense occlusion. Secondly, addressing the issue that semantic segmentation of daylilies requires complex topological structures, while traditional topological coding strategies have limited structural expressive power and cannot represent core topologies such as residual jump connections, and the existing general NAS search space is not adapted to the daylily scenario, this invention proposes a high-degree-of-freedom topological coding strategy that supports single-node multi-precedence node connections. This breaks through the constraint that traditional single-node connections can only connect to a single preceding node, effectively representing core topological structures such as residual jump connections and multi-branch parallelism, expanding the effective coverage of the search space, improving the ability to discover optimal architectures, and adapting to the segmentation needs of daylily scenarios where neck and leaves are difficult to distinguish and dense occlusion occurs. Third, addressing the issues of poor adaptability of evolutionary operators in traditional multi-objective NAS, which easily disrupt the cell structure and semantic integrity of daylily-related scenarios, and suffer from insufficient search efficiency and effective spatial coverage, making it unable to efficiently adapt to scenarios with dense occlusion and high neck-leaf similarity in daylilies, this invention designs a dual-mode crossover operator and a topology-operation cooperative mutation operator adapted to cell-level segmented encoding. This operator, while ensuring the integrity of the cell structure and semantic integrity adapted to daylily scenarios, effectively improves the search efficiency and effective spatial coverage of feature processing architectures adapted to dense occlusion and high neck-leaf similarity scenarios, facilitating the rapid discovery of high-quality daylily segmentation architectures. Fourth, addressing the problem that traditional congestion assessment methods lack adaptability, leading to search architectures that cannot meet core requirements such as the accuracy of distinguishing daylily necks and leaves and the effectiveness of occluded region segmentation, thus exacerbating solution set homogenization and affecting NAS optimization performance, this invention constructs an adaptive fusion congestion calculation mechanism between the target performance domain and the architecture structure domain. By dynamically allocating weights through dual-domain variance, it focuses on core performance requirements such as the accuracy of distinguishing daylily necks and leaves and the effectiveness of occluded region segmentation, while also considering the adaptability of the architecture to customized cell structures, effectively alleviating the solution set homogenization problem and ensuring that the searched architecture is adapted to the complex segmentation scenario of daylilies. Fifth, addressing the shortcomings of existing training-free NAS evaluation methods, such as their single evaluation dimension, simple nonlinear fusion, inability to comprehensively characterize model performance, and insufficient reliability of performance predictions, which restrict the efficiency of NAS optimization for daylily scenarios and fail to meet the segmentation requirements of densely occluded daylily plants and difficult-to-distinguish neck leaves, this invention proposes a four-dimensional collaborative low-cost neural architecture performance evaluation strategy. Using model perturbation robustness, feature space isotropy, parameter-gradient joint uncertainty, and gradient-layer structure regularity as core evaluation indicators, the candidate architecture is comprehensively evaluated. Each dimension accurately adapts to the needs of daylily harvesting scenarios, effectively overcoming the one-sidedness of traditional single-dimensional evaluation and providing precise guidance for optimizing daylily segmentation architectures. Sixth, addressing the significant limitations of traditional simple linear fusion of multidimensional evaluation metrics, which struggles to balance computational overhead and performance prediction reliability, thus failing to efficiently support NAS optimization in daylily scenarios and consequently affecting the segmentation accuracy and efficiency of daylily, this invention designs a ranking-based nonlinear geometric fusion mechanism. By scientifically integrating the scores of four-dimensional evaluation metrics, it improves the correlation between performance prediction accuracy and actual performance while strictly controlling computational overhead and achieving low-cost, rapid evaluation. This significantly shortens the NAS optimization iteration cycle and provides accurate and reliable guidance for multi-objective architecture search adapted to scenarios with dense occlusion and difficulty in distinguishing neck leaves in daylilies. Seventh, addressing the issues of insufficient robustness in the localization of daylily picking points, susceptibility to localization errors due to dense occlusion and high similarity of neck leaves, and the lack of localization algorithms adapted to this scenario, this invention proposes the MW-GCA intelligent localization algorithm for daylily picking points. Based on the optimal semantic segmentation model obtained through NAS optimization, it uses a "U"-shaped local window to filter corner information and a pose line localization method to achieve accurate estimation of the pixel coordinates and angles of the picking point. This algorithm is suitable for complex scenarios with dense occlusion and high similarity of neck leaves, thus solving the pain point of inaccurate localization of daylily picking points. Eighth, in the automated harvesting of daylilies, the information of the harvesting point cannot be effectively converted into executable instructions for the robotic arm, and the plants are easily damaged and their appearance is ruined due to positioning deviations and improper instruction conversion. In addition, the positioning problem caused by the dense shading of daylilies and the difficulty in distinguishing the neck leaves is also addressed. This invention uses a three-dimensional spatial coordinate transformation algorithm and a robotic arm kinematics analysis algorithm to convert the four-dimensional harvesting information obtained by the MW-GCA algorithm into robotic arm action instructions, so as to complete the automated and precise harvesting of daylilies, reduce the plant damage rate, maintain the integrity of the appearance, and improve the economic benefits of the industry. This invention features a rationally designed technical architecture that specifically addresses the core technical bottlenecks in NAS-based semantic segmentation and automated harvesting of daylilies. By automating manual labor, it significantly reduces manpower input and improves operational efficiency, aligning with the national strategy of intelligent manufacturing and independent innovation. This will accelerate the intelligent transformation of the daylily industry and promote its sustainable development. Given its practicality and industrial empowerment value, this invention possesses broad market application space and commercialization potential. Attached Figure Description

[0005] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. To clearly present the technical solutions of the embodiments of the present invention, the accompanying drawings involved in the description of the embodiments will be briefly described below: The accompanying drawings are used to intuitively and clearly illustrate the implementation methods, overall architecture, and core concepts of the present invention, assisting the specification in explaining the principles of the present invention, and helping those skilled in the art to quickly understand the technical solutions and application logic of the present invention. For those skilled in the art, based on the above-described accompanying drawings, other related drawings can be reasonably derived without creative effort; Figure 1 This is a general framework diagram of the present invention; Figure 2 Candidate architecture structure diagram; Figure 3 Cell structure diagram for each stage; Figure 4 This is a diagram of the FPA cell structure. Figure 5 Here is a flowchart of the search strategy; Figure 6 Flowchart of the MW-GCA picking point location algorithm; Figure 7 Flowchart of the algorithm for determining the daylily cutting angle based on Mini-Window; Figure 8 Scatter plots showing the correlation between the model's actual test accuracy and predicted score on different datasets; Figure 9 Radar plots for evaluating the correlation of methods under different datasets and sample sizes; Figure 10 This is a picture showing the actual harvesting effect of the robotic arm. Detailed Implementation

[0006] To better understand the purpose, technical solution, and beneficial effects of this invention, the following description is provided in conjunction with the appendix. Figure 1-10 Specific embodiments of the present invention are described in detail herein to further explain the invention, rather than to limit the scope of protection of the invention. Those skilled in the art should understand that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. The overall framework of the automatic architecture method for the H²D²-MONAS daylily segmentation model is as follows: Figure 1 As shown, the framework clearly presents the collaborative mechanism of the core modules within the H²D²-MONAS method. This method takes a daylily segmentation dataset and a customized search space as input. Through the collaborative optimization of three core modules, it ultimately outputs the optimal semantic segmentation model adapted to the daylily harvesting scenario. Specifically, the hierarchical heterogeneous multi-source fusion search space for daylily segmentation defines the search boundary for scenario adaptation; the multi-objective neural architecture search strategy guided by dual-domain semantic collaboration achieves efficient global optimization of candidate architectures; and the low-cost neural architecture performance evaluation strategy based on multi-dimensional collaboration and ranking provides accurate and low-overhead performance support for architecture selection. Based on this, the MW-GCA algorithm accurately locates the four-dimensional information of the harvesting point, ultimately enabling automated daylily harvesting through collaboration with a robotic arm execution system. The following sections will elaborate on the specific implementation process and operational details of each core component.

[0007] First, a heterogeneous, staged search unit for the encoder-decoder end-to-end for daylily segmentation tasks. To address the core issues of severe occlusion, high similarity between neck leaves and stem leaves, and difficulty in classifying daylily buds during harvesting, particularly the challenges of feature loss due to dense occlusion and the difficulty in fusing multi-scale features of neck leaves, leading to misclassification, this invention focuses on the practical needs of daylily semantic segmentation. It constructs a hierarchical heterogeneous cell structure adapted to the daylily harvesting scenario and specifically designs a phased Normal / Attention adaptation unit and a feature pyramid aggregation (FPA) decoding unit for multi-scale fusion within the NAS search space. The phased Normal / Attention adaptation unit effectively preserves features and suppresses field interference in densely occluded daylily bud scenarios. The FPA decoding unit, through hierarchical aggregation of multi-scale features, supplements missing features in occluded areas and enhances the discriminative power of different scale features in neck leaves. This addresses the core pain points of daylily harvesting, improves the accuracy of daylily semantic segmentation, and provides reliable structural support for daylily segmentation tasks. Specifically, the overall framework of the NAS candidate architecture for daylily segmentation designed in this invention is as follows: Figure 2 As shown, the input image of daylily in a field is first preprocessed by two 3×3 convolutional modules, with a stride of 2 for each convolution to halve the size. After two layers of convolution, the number of channels in the feature map is expanded to 32, completing the initial feature extraction and dimension regularization. Subsequently, the preprocessed feature map is sequentially fed into a three-level feature extraction unit for depth mining. Each level unit consists of a downsampling unit (ResCell), two deep feature extraction modules (FECell), and an attention module (AttCell) cascaded together. After each level of processing, the feature map size is halved and the number of channels is doubled. Through three-level progressive extraction, the deep fine-grained features under dense occlusion are fully mined, and the subtle differences in the neck and leaves are accurately captured. Finally, the multi-scale feature map output by the three-level unit is fed into the HFPA hierarchical feature pyramid aggregation decoding module. After passing through FPACell S1 and FPACell S2 units, 1×1 convolution, and four times bilinear upsampling, the original size of the input image is restored, and the final semantic segmentation result is output. The three-level feature extraction units ResCell, FECell, and AttCell have the following structures respectively: Figure 3As shown in (a), (b), and (c), all designs were specifically tailored to the segmentation requirements of daylily. ResCell contains two computation nodes, and the optional operations for each node are shown in Table 1. The input features are processed in parallel by the two nodes and then output by channel splicing. This can effectively preserve the basic structural features of the plant while achieving downsampling and channel expansion, and avoid the loss of neck leaf features. FECell consists of 5 nodes, with x0 as the input node and x6 as the output node. The optional operations for the intermediate nodes x1 to x5 are shown in Table 2. Each node x1 to x5 must be connected to at least one preceding node. The output node x6 splices the features of all intermediate nodes without subsequent connections, which can enhance the expression of subtle differences in neck leaf features and solve the problem of difficulty in distinguishing neck leaf categories. AttCell contains 4 computation nodes, and the optional operations for each node correspond to spatial, channel, edge, and texture attention, respectively. The input features are processed in parallel by the four nodes and then spliced ​​to generate attention weights. These weights are then multiplied element-wise with the preceding feature map, which can suppress invalid background information in the field and highlight key features of daylily, thereby improving the robustness of the model to complex field disturbances. Table 1 ResCell Candidate Operation Table Serial Number Candidate operations 0 3×3 depthwise separable convolution 1 3×3 max pooling 2 3×3 average pooling Table 2 FECell Candidate Operation Table Serial Number Candidate operations Serial Number Candidate operations 0 Equality 4 3×3 average pooling 1 3×3 depthwise separable convolution 5 1×3+3×1 convolution 2 5×5 depthwise separable convolution 6 1×5+5×1 convolution 3 3×3 max pooling 7 1×7+7×1 convolution This invention addresses the multi-scale feature fusion requirements of daylilies by designing two types of differentiated decoding units, FPACell S1 and FPACell S2, which are applied to different levels of decoding, as shown in the structure below. Figure 4 As shown in (a) and (b), FPACell S1 is the first-level decoding unit, which directly fuses the multi-scale features extracted from the three-level feature extraction. It includes two input nodes, three intermediate nodes, and one output node. The node optional operations are shown in Table 3. The x2 node is upsampled by 2 times to unify the size. After channel splicing and 1×1 convolution to normalize the number of channels, the second-level and third-level features are fused to complete the primary multi-scale feature decoding, which can achieve effective localization of plants at all scales. FPACell S2 is the second-level decoding unit, which inherits the results of FPACell S1 and performs deep fusion. It includes two input nodes and one output node. Similarly, after upsampling by 2 times to unify the size, channel splicing and 1×1 convolution to normalize the number of channels, the output of FPACell S1 and the first-level features are fused. This can complete the features of the occluded area and ensure the precision and completeness of the daylily segmentation results. Table 3 Candidate Operation Table for AttCell and FPACell Serial Number Candidate operations Serial Number Candidate operations 0 Equality 6 5×5 depthwise separable convolution 1 3×3 max pooling 7 7×7 depthwise separable convolution 2 3×3 average pooling 8 3×3 depthwise separable convolution (dilation rate of 2) 3 5x5 max pooling 9 3×3 depthwise separable convolution (dilation rate of 4) 4 5×5 average pooling 10 5×5 depthwise separable convolution (dilation rate of 2) 5 3×3 depthwise separable convolution 11 5×5 depthwise separable convolution (dilation rate of 5)

[0008] Second, it supports a high-degree-of-freedom topology coding strategy that supports single-node multi-precedence node connections. To address the limitations of traditional topological coding strategies, which typically rely on single-node connections to only a single preceding node, severely restricting structural expressiveness and failing to represent the complex topological structures required for daylily harvesting scenarios characterized by dense occlusion, similar neck leaf textures, and the coexistence of multiple scale targets, this invention proposes a high-degree-of-freedom topological coding strategy that supports single-node connections to multiple preceding nodes. This strategy is also compatible with the aforementioned hierarchical heterogeneous cell structure. By overcoming the constraints of traditional single preceding connection, it can encode residual skip connections to preserve the subtle textures of daylily buds and neck leaves, implement multi-branch parallel processing to collaboratively handle multiple heterogeneous features, and perform cross-level, multi-scale fusion to supplement missing features in mutually occluded areas. This expands the effective coverage of the search space and enhances the ability to discover optimal architectures. Specifically, the high-degree-of-freedom topological coding strategy proposed in this invention is specifically adapted to the aforementioned hierarchical heterogeneous Cell structure (ResCell, FECell, AttCell) and decoding unit (FPACell S1, FPACell S2). Through differentiated unit coding design, the innovative concept of connecting multiple preceding nodes in a single node is implemented, achieving accurate representation of the complex topological structure required for semantic segmentation of daylilies, and providing core coding support for discovering the optimal semantic segmentation architecture adapted to the picking scenario. For the coding design of the feature extraction stage (Stage part), this invention combines the functional characteristics of each Cell unit with the specific requirements of daylily segmentation and designs a differentiated coding scheme: the ResCell coding consists of 2 bits, corresponding to the candidate operation numbers of its two nodes respectively. The coding is concise and accurately matches its core functions of downsampling and basic texture preservation, providing support for the subsequent basic feature differentiation of daylily buds, neck leaves, and background; for the core positioning of FECell, which is responsible for the collaborative extraction of multi-branch features in densely occluded areas and solving the pain point of mutual occlusion of daylily buds, this invention uses a combination of "operation coding + connection coding". The odd-numbered bits are operation codes, corresponding to the candidate operations of each node, and the even-numbered bits are connection codes, representing the connection relationship between each node and its predecessor node. The decimal value of the connection code corresponds to the binary form, and the number of binary bits is consistent with the node number. Taking the x3 node as an example, its connection code corresponds to a 3-bit binary number, which, from left to right, represents the connection status with the 0, 1, and 2 predecessor nodes respectively (1 indicates connected, 0 indicates not connected). Figure 3As shown in (b), when the connection encoding is 7, the binary form is 111, meaning that node x3 is connected to all preceding nodes 0, 1, and 2. This design can flexibly realize arbitrary connections between a single node and multiple preceding nodes, adapting to the requirements of FECell multi-branch feature collaborative extraction, and helping to uncover subtle boundary features of daylily neck leaves and buds under dense occlusion. The overall encoding length of FECell is 10 bits. The encoding of AttCell consists of 4 bits, corresponding to the candidate operation numbers of its four nodes, accurately adapting to its core function of attention weight generation, and can encode the feature enhancement of key areas of daylily buds, suppressing redundant interference from the background and neck leaves. The encodings of ResCell, FECell, and AttCell (a total of 16 bits) together constitute the encoding of a single feature extraction stage. The encodings of the three feature extraction stages total 48 bits, forming the complete encoding of the feature extraction backbone network, realizing accurate representation of the deep feature extraction topology. For the encoding design in the decoding stage, combining the differentiated decoding functions of FPACell S1 and FPACell S2 with the segmentation requirements of daylily, an encoding scheme of corresponding length was designed: FPACell S1, as the first-level decoding unit, has an encoding length of 6 bits, corresponding to the candidate operations of each node, adapting to the requirements of primary fusion of multi-scale features, and can encode the initial alignment of daylily buds at different scales and feature completion of occluded areas; FPACell S2, as the second-level decoding unit, has an encoding length of 3 bits, corresponding to the candidate operations of each node, adapting to the requirements of deep multi-scale feature fusion, and can encode the fine distinction of the boundary features of buds and neck leaves, solving the missegmentation problem caused by the high similarity of their textures. Overall, the architecture encoding of this invention consists of 48-bit encoding for three feature extraction stages, 6-bit encoding for two FPACell S1 stages, and 3-bit encoding for one FPACell S2 stage, totaling 63 bits, constructing an optimal search space encoding system adapted to the daylily harvesting scenario. This system, through highly flexible connection representation and differentiated unit encoding design, effectively overcomes the structural limitations of traditional topological encoding. It can accurately represent topological structures such as residual jump connections, multi-branch parallelism, and cross-level multi-scale fusion, comprehensively covering the core scenario requirements of dense occlusion, similar neck leaf texture, and coexistence of multi-scale targets in daylilies. This further expands the effective coverage of the search space, enhances the mining capability of the optimal semantic segmentation architecture, and provides core encoding support for achieving synergistic optimization of semantic segmentation accuracy and inference efficiency in daylily harvesting.

[0009] Third, the dual-mode crossover operator and the topology-operation cooperative mutation operator. To address the issues that traditional multi-objective NAS general evolutionary operators do not consider the specific feature extraction requirements for daylily semantic segmentation, and that crossover mutation easily disrupts the cell structure and coding semantic integrity designed for the overlapping neck and leaf characteristics and dense occlusion of daylily fruits, as well as insufficient search efficiency and effective spatial coverage, this invention designs a dual-mode crossover operator and a topology-operation cooperative mutation operator adapted to cell-level segmented coding. This operator, by adapting to cell-level segmented coding, effectively improves the search efficiency and effective spatial coverage of the feature extraction architecture for small target fruits under conditions of blurred fruit and neck leaf boundaries and dense occlusion, while ensuring the integrity of the cell structure and coding semantics optimized for daylily feature extraction. This helps to quickly discover high-quality architectures, ultimately achieving a synergistic improvement in global optimization performance, search efficiency, and solution set diversity. This provides a new technical path for high-precision, lightweight semantic segmentation architecture search for different deployment constraints in daylily harvesting scenarios. like Figure 5 As shown, the core process of the dual-domain semantic collaborative guided multi-objective neural architecture search strategy proposed in this invention strictly follows the following steps: After the algorithm starts, the population is first initialized using the optimal point set algorithm to provide a uniformly distributed initial candidate architecture for the high-dimensional search space, avoiding the initial population from being concentrated in invalid regions unsuitable for daylily feature extraction; then, the performance of the initial population is evaluated, and non-dominated sorting and dual-domain adaptive fusion crowding calculation are completed; next, it is determined whether the iteration termination condition is met. If it is met, the Pareto optimal solution set is directly output and the process ends; if it is not met, the evolutionary iteration stage is entered: the dual-domain crowding tournament selection is executed sequentially to screen high-quality parent individuals, and the dual-mode crossover operator of cell-level exchange and uniform crossover is completed. The architecture is re-encoded and reorganized, and then the topology and operation branches are co-perturbed by the topology-operation cooperative mutation operator. The whole process strictly adapts to the cell-level segmented encoding and performs legality repair to ensure the integrity of the cell structure and encoding semantics optimized for daylily feature extraction. After the evolutionary operation is completed, the performance of the generated offspring individuals is evaluated. Then, the parent and offspring individuals are merged, and the new generation population is selected through non-dominated sorting, dual-domain adaptive fusion crowding calculation and elite retention strategy. Then, the termination condition is checked to enter the next round of iteration. The iteration is repeated until the termination condition is met, and the Pareto optimal architecture solution set that balances the number of parameters and the segmentation performance is output to complete the search for a high-precision and lightweight semantic segmentation architecture for daylily harvesting. Specifically, the dual-mode hybrid crossover operator designed in this invention is specifically adapted to the characteristics of cell-level segmented encoding. Addressing the shortcomings of traditional general crossover operators, which easily disrupt the semantic integrity of cell-level encoding optimized for daylily feature extraction and fail to simultaneously preserve the inheritance of high-quality daylily-specific cell structures and the expansion of fine-grained search space, this invention fuses cell-level exchange crossover and uniform crossover modes with a total crossover probability of 0.9 and a probability weight of 0.6:0.4. This achieves efficient reorganization of the architecture encoding while ensuring the integrity of the cell structure. Specifically, the cell-level exchange crossover is performed with a probability of 0.6, randomly selecting a complete cell unit and exchanging all the corresponding encoding bits of that cell in the two parent candidate architectures, completely preserving the parent architecture. The structure and semantics of high-quality cells have been verified to be effective in extracting edge features of daylily fruits and distinguishing between neck leaves and fruits. This avoids destroying the topology of feature extraction units optimized for dense occlusion and small target fruit localization in daylilies. It quickly inherits the ability of high-quality architecture to segment overlapping fruits and distinguish neck leaves from fruits. Uniform crossover is performed with a probability of 0.4. For each bit of the 63-bit architecture encoding, the corresponding bit values ​​of the two parents are swapped with a probability of 0.5 to achieve fine-grained architecture feature perturbation. Without destroying the overall cell structure, it expands the effective coverage of the search space for feature extraction under different light and occlusion levels of daylilies and explores potential high-quality architecture topologies that are more adapted to complex field environments. To address the problem that traditional mutation operators separate topology mutation from operation branch mutation, which can easily lead to topology-operation semantic mismatch and network structure collapse in daylily feature extraction, this invention proposes a topology-operation cooperative mutation operator. This operator binds topology connections and corresponding operations into a unified semantic unit for cooperative mutation, ensuring the coordination of mutations and the effectiveness of the architecture in the daylily feature extraction task. This operator performs differentiated cooperative mutation on topology connection bits and operation bits in cell-level encoding. On the one hand, it achieves topology-operation cooperative perturbation. For valid connection bits in the encoding, an existing connection is randomly selected, and two associated operations are executed simultaneously: first, the operation label of the node corresponding to the connection is modified with a preset probability, and a new operation is randomly selected from the candidate operation set; second, the connection is deleted with a preset probability, and a completely new valid connection is randomly added and assigned a random operation. Through this cooperative operation of deleting the old and adding the new, it avoids the problem of insufficient node input caused by only performing deletion. This approach addresses the issue of network structure collapse by forcing the exploration of new connection-operation combinations that are more suitable for multi-scale feature fusion and small-target plant enhancement in daylilies, thereby improving the targeting and quality of mutations. Furthermore, it achieves fine-grained mutation of operation branches by randomly updating the operation type within the candidate operation set for the operation bits in the encoding, thus optimizing the operation branches to meet the feature extraction needs of different levels, such as shallow texture feature extraction and deep semantic feature fusion in daylilies. Simultaneously, the operator strictly adheres to the rigid constraints of cell-level encoding, locking the non-mutable core bits and performing legality checks and repairs after mutation to ensure that all encoding bits conform to boundary constraints and semantic rules, guaranteeing the integrity of the cell structure optimized for daylily feature extraction. This fundamentally solves the problem of traditional mutation operators easily generating invalid architectures and causing network structure collapse, while also strengthening the exploration capability of new connection-operation combinations and improving the efficiency and effectiveness of architecture search.

[0010] Fourth, a dual-domain adaptive fusion congestion calculation method. To address the core bottlenecks of traditional multi-objective Neural Architecture Search (NAS), such as insufficient adaptability and single-dimensional crowding assessment, which easily leads to homogenized solution sets and a search architecture biased towards generalization and lightweight approaches, resulting in the loss of segmentation capabilities for densely occluded flower buds and the ability to distinguish the boundaries between neck leaves and fruits in daylilies, as well as low global optimization and search efficiency, this invention proposes a dual-domain adaptive fusion crowding calculation method. It constructs an adaptive fusion mechanism between the target performance domain and the architecture structure domain. The target performance domain balances the model evaluation score with its own complexity through a low-cost evaluation strategy, focusing on the core indicators of daylily semantic segmentation. The architecture structure domain corresponds to the cellular structure features customized for daylily feature extraction in this invention. By dynamically allocating weights through dual-domain variance, it balances the core performance of daylily semantic segmentation with the adaptability of the architecture to the customized cellular structure, effectively alleviating the problem of homogenized solution sets, avoiding the pain point of achieving general performance but insufficient daylily segmentation accuracy, improving global optimization performance and search efficiency, and perfecting the multi-objective neural architecture search system for this scenario. Specifically, this invention addresses the semantic segmentation requirements and hierarchical heterogeneous cell structure characteristics of daylily harvesting scenarios. It constructs a dual-domain congestion evaluation system comprised of the target performance domain and the architectural structure domain. The final congestion score is obtained through adaptive weighted fusion, with the following fusion formula: ; In the formula, For the final integration of individual crowding, For the target domain congestion, For the congestion of the structural domain, The weights are dynamically adaptive, and the crowding levels of both domains are normalized to the [0,1] interval. The larger the value, the more crowded the individual distribution under the corresponding dimension. Individuals with lower crowding levels are selected first during the screening process. Target domain crowding is used to measure the density of candidate architectures in the optimization target space. The optimization target includes two metrics: the number of model parameters and the low-cost evaluation score. The calculation first performs min-max normalization on the target values ​​of individuals within the same non-dominated front to eliminate dimensional differences. Then, for each individual, the two nearest neighbors in the target space with the closest Euclidean distance are selected, and the original target domain crowding is calculated based on the nearest neighbor distance, as shown in the following formula: ; In the formula, For individuals i and neighbors j In the Euclidean distance of the target space, To prevent the use of tiny constants with a denominator of zero, for individuals located at the boundary of the target space, their original crowding degree is directly set to 0. Then, the original target domain crowding degree of all individuals is normalized using a min-max method to obtain the normalized target domain crowding degree. ; Domain crowding is used to measure the similarity of candidate architectures in hierarchical heterogeneous cell topologies. This invention uses the 63-dimensional encoding vectors corresponding to each architecture as the basis for structural representation to achieve a digital description of the topology. Subsequently, the cosine similarity between the encoding vectors of different individuals is calculated. For each individual, the two nearest neighbors with the highest similarity are selected, and the original domain crowding is calculated based on structural similarity. The calculation formula is as follows: ; In the formula, For individuals i and neighbors j The structural cosine similarity is calculated. After the original crowding degree is calculated, the results of individuals within the same front are normalized using the min-max method to obtain the normalized domain crowding degree. ; To achieve dynamic adaptive allocation of weights, this invention calculates the fusion weights based on the distribution variance of the two-domain crowding degree within the current non-dominated frontier. The calculation formula is as follows: ; In the formula, and These represent the variances of the target domain crowding and the structural domain crowding within the current frontier, respectively. These weights can be automatically adjusted based on population distribution. When the structural domain distribution is more dispersed, the target domain weight is increased to ensure diversity in segmentation performance; when the target domain distribution is more dispersed, the structural domain weight is increased to maintain architectural topological diversity and avoid excessive homogeneity of solution sets. This dual-domain adaptive fusion crowding calculation method is integrated into the survival selection stage of the improved NSGA-II algorithm. After merging the parent and offspring populations, it first divides the individuals into hierarchical levels using Pareto non-dominated sorting, then calculates the dual-domain crowding at each level and completes adaptive fusion. Finally, it selects elite individuals according to the rule of prioritizing non-dominated levels and ascending crowding at the same level. This avoids the problem of retaining a large number of homogeneous individuals with similar performance but weak ability to extract daylily-specific features when traditional methods only select based on general performance crowding. By taking into account both daylily-specific performance indicators and customized cellular structure features in a dual-dimensional evaluation, this method effectively solves the problems of single-dimensional crowding evaluation and poor scene adaptability in traditional methods. It significantly improves the diversity of the population in extracting daylily-specific features and the global search efficiency, providing a reliable evaluation basis for accurately mining lightweight semantic segmentation architectures that take into account both robustness in complex field environments and deployment requirements of edge harvesting equipment.

[0011] Fifth, multi-dimensional low-cost model performance evaluation strategy To address the core bottlenecks of existing training-free NAS evaluation methods, such as their single evaluation dimension, simple nonlinear fusion, inability to comprehensively characterize the model's overall performance in the daylily scenario, insufficient reliability of performance prediction, and limitations on global optimization efficiency, this invention proposes a low-cost neural architecture performance evaluation strategy based on multi-dimensional collaboration. This strategy constructs an efficient NAS performance evaluation system adapted to the daylily harvesting scenario, with the core objective of improving NAS optimization efficiency. The principle is to use model perturbation robustness, feature space isotropy, parameter-gradient joint uncertainty, and gradient-layer structure regularity as core evaluation indicators to conduct a comprehensive evaluation of candidate architectures without training. Each dimension is specifically adapted to the core needs of the daylily scenario: model perturbation robustness ensures segmentation stability under field light fluctuations, leaf swaying, and different growth stages; feature space isotropy enhances the ability to identify buds facing different directions under dense occlusion and distinguish subtle boundaries between leaf necks and buds; parameter-gradient joint uncertainty improves the generalization performance of small target bud segmentation; and gradient-layer structure regularity adapts to the lightweight deployment requirements of edge-harvesting equipment, overcoming the one-sidedness of traditional evaluations based solely on model complexity or single validation accuracy. This strategy significantly improves the performance prediction reliability of NAS in the daylily scenario, solves the problem of traditional evaluation methods restricting optimization efficiency, provides reliable evaluation support for efficient optimization of NAS in the daylily harvesting scenario, and helps to quickly discover high-quality lightweight semantic segmentation architectures adapted to complex field environments. Specifically, this invention uses model perturbation robustness as the first core evaluation dimension to quantify the stability and robustness of candidate architectures when faced with input perturbations in real field environments. The specific evaluation process is as follows: First, addressing the common problems of image information loss and partial occlusion in daylily harvesting scenarios, this invention perturbs the original input image through a center cropping operation, generating interfering test input samples. Its generation formula is defined as follows: ; in, Center-based clipping function; Image scaling function; input tensor This represents a single batch of input images, where dimensions B, C, H, and W represent the batch size, number of channels, image height, and width, respectively. During the evaluation process, this invention uses batches containing 128 samples as input to ensure that the evaluation results fully characterize the overall distribution features of the dataset. Subsequently, to eliminate the dimensional differences between different feature dimensions and improve sensitivity to subtle changes in feature distribution, this invention uses the rank correlation coefficient principle to measure distribution consistency. Specifically, the feature vectors extracted from the original input sample and the perturbed sample by the model are sorted by element value and assigned corresponding rank coefficients, thereby constructing the rank matrix corresponding to the original input. Rank matrix corresponding to the perturbation input Based on this, the present invention calculates the average rank correlation coefficient of a single batch of samples. To quantify the distributional consistency between two-rank matrices, the calculation formula is as follows: ; In the formula, and Let be the mean of the rank vector of the b-th sample. To prevent the smoothing constant from having a denominator of zero; Finally, to map the evaluation results to a unified quantization range, this invention introduces an exponential mapping function to transform the rank correlation coefficient into a robust evaluation score in the range of 0 to 100. The specific formula is defined as follows: ; The higher the score, the more stable the feature distribution of the candidate architecture is when faced with input perturbations, and the stronger the model's robustness to perturbations. Through this dimension of evaluation, the present invention effectively quantifies the deployment reliability of candidate architectures in real-world daylily harvesting scenarios such as dense occlusion and information loss, providing key quantitative basis for screening high-quality architectures with high environmental robustness; This invention uses the isotropy of the model feature space as the second core evaluation dimension to quantify the feature representation quality and information redundancy of candidate architectures, thereby selecting high-quality architectures that are suitable for distinguishing subtle features of daylily neck leaves. The specific evaluation process is as follows: Specifically, the single-batch input image sampling strategy consistent with the model perturbation robustness assessment is used, and the input tensor... Each training batch consists of 128 images randomly sampled from the dataset, ensuring that the evaluation results fully reflect the overall distribution characteristics of the dataset. It relies on model forward propagation and hook registration to collect images from the network's front... L The output feature tensor of the first main block; for the first... l Each main block has output characteristics ,in For the number of channels, and Represent the height and width of the feature map, and reshape the feature tensor into a feature matrix. ,in The total number of feature samples; Next, the characteristic matrix The centered feature matrix is ​​obtained by performing a centralization process. Then add regularization parameters. Calculate the regularized covariance matrix to avoid matrix singularity issues caused by insufficient samples. Based on this, calculate the real symmetric covariance matrix. Perform eigenvalue decomposition to obtain eigenvalues. Then set a threshold. By performing nonnegative adjustment and L1 normalization on the eigenvalues, the principal component probability distribution is obtained. The specific formula is defined as follows: ; Furthermore, the probability distribution of the principal components is used to calculate the first... l Feature representation entropy of each main block To quantify the dispersion of the principal block's feature distribution, the specific formula is defined as follows: ; Finally, all L The feature representation entropy of each principal block is accumulated to obtain the overall feature space isotropic evaluation score of the architecture. The specific formula is defined as follows: ; The higher the score, the more uniform the feature distribution of each principal block of the model, the more significant the isotropy of the feature space, and the lower the feature expression redundancy. It can effectively enhance the ability to distinguish the subtle features of the neck and leaves of daylilies, adapt to the multi-scale feature extraction needs in dense occlusion scenarios, make up for the shortcomings of traditional single-dimensional evaluation in characterizing the quality of feature expression, and provide core quantitative support for the feature expression level of multi-dimensional collaborative evaluation system. This invention uses parameter-gradient joint uncertainty as the third core evaluation dimension to quantify the parameter information density and generalization performance of candidate architectures, thereby screening high-quality architectures that meet the generalization requirements of daylily harvesting scenarios. The specific evaluation process is as follows: Specifically, the multi-factor weighting strategy proposed in this invention uses a two-dimensional matrix flattened from the parameters of each layer. W Based on the computational foundation, the comprehensive importance weight of each layer is quantified by integrating key weighting factors from four dimensions. Among them, the basic scale weight is defined as... , For parameter matrix W The total number of elements, which characterizes the computational complexity of the model through the total number of parameters, is a fundamental dimension for importance assessment; the network depth coefficient is defined as... ,in b This is the index of the main block where the layer is located. B The total number of main blocks is used to model the position of the layers in the gradient propagation path. The deeper the network layer, the more significant its contribution to the model output, adapting to the deep network requirements for daylily feature extraction; the module stacking coefficient is defined as... , The number of normal unit stackings is used to characterize the feature accumulation effect when modules are stacked. The more stackings, the higher the weight, adapting to the stacking structure requirements of multi-scale feature extraction; functional role weight. The contribution of different layers to the network function is assigned values: the downsampling layer, responsible for preserving discriminative information and significantly impacting the final model's performance and efficiency, is assigned the highest weight of 1.60; the feature extraction layer, primarily tasked with extracting subtle features from the daylily's neck and leaves, is assigned a weight of 1.35, higher than the baseline; the preprocessing layer, used to ensure the stability of the input data, is assigned a weight of 1.10, slightly higher than the baseline; the classification layer, heavily reliant on preceding feature outputs, has a relatively low weight, set at 0.90; the global pooling layer and other layers serve only as bridges for information exchange, and are assigned the baseline weight of 1.00. Finally, the four weight factors are weighted and fused using the following formula to obtain the comprehensive importance weight of each layer: ; For weight matrices of different scales, this invention first performs singular value decomposition on them to obtain singular value sequences. Then, a cumulative energy threshold was set. effective rank Defined as satisfying The smallest positive integer, where Let be the full rank of the matrix; if the matrix norm This indicates that the information contribution of this layer is weak during the initialization phase, and the effective rank ratio of this layer is set to be... Otherwise, the effective rank ratio is defined as Finally, this invention uses a comprehensive importance weight for each layer. I As a weighting factor, the rank ratio of the effective layer is... A weighted average is then performed to obtain a score for the joint uncertainty of the model parameters and gradients. The specific mathematical expression is as follows: ; The higher the score, the higher the density of model parameter information and the stronger the generalization performance. It can effectively adapt to the generalization needs of complex field environments in daylily harvesting scenarios, improve the segmentation stability of the model under different lighting conditions, dense shading and other conditions, make up for the shortcomings of traditional single-dimensional evaluation in characterizing generalization performance, and provide core quantitative support for the generalization performance level of multi-dimensional collaborative evaluation system. This invention uses gradient-layer structure regularity as the fourth core evaluation dimension to quantify the gradient distribution coordination and structural regularity of candidate architectures, thereby screening high-quality architectures that meet the lightweight deployment requirements of daylily harvesting scenarios. The specific evaluation process is as follows: Specifically, to obtain the statistical features of the model more stably, this invention generates... N Group(N =5) Input samples conforming to a standard normal distribution and their corresponding target labels are used to generate data that is input into the model to be evaluated to perform backpropagation, thereby obtaining the gradients of the trainable parameters in the candidate architecture. ( L (The loss value of the loss function). For different parameters W and its gradient First, calculate its direction perception uncertainty. The specific formula is defined as follows: ; In the formula, Indicates the learning rate. It is a smoothing factor; Based on this, N Valid data obtained from the second sampling Merge into vectors (Remove items with absolute values ​​less than) (approximately zero value), calculate the mean of its absolute value sequence. and standard deviation .like Then, the absolute value of each element is obtained through a Gaussian distribution. Fit the probability distribution; if This indicates a highly concentrated distribution, and a uniform distribution is used for fitting. Subsequently, the fitted probability distribution is... Calculate the gradient entropy of this layer. The specific formula is defined as follows: ; Furthermore, to further quantify the differentiated impact of different layer types on the model training process and to adapt to the hierarchical structure characteristics of the daylily semantic segmentation network, this invention introduces a role multiplier. (Numerical and parameter-gradient joint uncertainty dimensions are consistent), depth scaling (in b This refers to the index of the main block where the layer is located. B (Total number of main blocks) and stacking ratio (in s (The number of stacked Normal blocks), for different layers in the model. We perform weighting to obtain the weighted hierarchical entropy value. The specific formula is defined as follows: ; Finally, for all effective layers ( The total weighted entropy is obtained by summing the weighted entropies of units with length ≥ 10. And further calculate the average weighted entropy (in M(This represents the number of effective layers). If the value of M is 0, the score is returned directly. Otherwise, calculate the intermediate score S, and then normalize the intermediate score to the interval [0,100], which will be used as the final model gradient-layer structure regularity score. The formula for calculating the median score S is defined as follows: ; The higher the score, the more stable the gradient distribution of each layer of the model, the stronger the structural regularity, and the better the training stability. It can effectively adapt to the lightweight deployment requirements of edge harvesting equipment in daylily picking scenarios, reduce the difficulty of model training, improve the operational stability of the architecture in actual field environment deployment, make up for the shortcomings of traditional single-dimensional evaluation in characterizing structural regularity, provide core quantitative support for the structural regularity level of multi-dimensional collaborative evaluation system, and help to discover high-quality lightweight semantic segmentation architectures that are suitable for daylily picking scenarios.

[0012] Sixth, a ranking-based nonlinear geometric fusion mechanism To address the limitations of traditional multidimensional evaluation metrics using simple linear fusion, which struggles to balance segmentation accuracy and deployment overhead in the daylily scenario, leading to performance prediction biased towards a general, lightweight architecture and neglecting core harvesting requirements such as densely occluded bud segmentation and neck-bud boundary differentiation, resulting in insufficient prediction reliability and long NAS optimization iteration cycles, this invention introduces a ranking-based nonlinear geometric fusion mechanism. This mechanism scientifically integrates the scores of the aforementioned four-dimensional evaluation metrics, abandoning the shortcomings of linear fusion. While strictly controlling evaluation overhead and achieving low-cost, rapid evaluation, it achieves reasonable fusion of multidimensional metrics, effectively improving the model's performance prediction accuracy in the daylily scenario, enhancing the correlation and reliability between prediction results and actual harvesting performance, shortening the NAS optimization iteration cycle, providing accurate evaluation guidance for multi-objective architecture search, and coordinating with relevant search strategies to achieve a comprehensive improvement in NAS global optimization performance, search efficiency, and solution set diversity, efficiently discovering high-quality semantic segmentation architectures adapted to the daylily harvesting scenario. Specifically, targeting m Given one candidate architecture to be evaluated and four proxy evaluation metrics, this invention first integrates the original scores of each architecture under each metric into an m×4 dimensional score matrix. S Each element in the matrix corresponds to the original score of a single candidate architecture under a single evaluation metric. This is used to construct a multi-dimensional evaluation dataset that fully covers the field robustness, feature differentiation, cross-environment generalization performance, and edge deployment adaptability of the model in the daylily scenario, providing comprehensive and objective basic data for subsequent fusion computing. Subsequently, regarding the first j The present invention arranges the score vectors of each evaluation indicator in ascending order using a dense ranking method, and assigns a corresponding ascending rank to each element. Then, the rank ratio is obtained through normalization. This forms an m×4 dimensional rank ratio matrix. P This dense ranking process can effectively eliminate the differences in the dimensions between different evaluation indicators, fully preserve the relative merits of candidate architectures under each indicator, fundamentally avoid the problems of unreasonable manual weight setting and unbalanced indicator weights in linear fusion, and eliminate the bias of sacrificing dense occlusion segmentation accuracy due to manual bias towards lightweighting or ignoring robustness, which leads to a sharp drop in field deployment performance. It ensures the objectivity and fairness of the evaluation results in each dimension and provides a reliable relative performance basis for subsequent fusion calculations. Finally, for the first i The rank-ratio row vectors of the candidate architectures are used to calculate the final fusion score by taking the fourth root of the product of their elements. The specific formula is defined as follows: ; This fusion method is a nonlinear geometric fusion, which can fully retain the relative advantages and disadvantages of each dimension of evaluation, accurately identify candidate architectures that perform evenly across the core requirements of daylilies, and avoid the misselection of architectures that are extremely dominant in a single dimension but have poor actual harvesting results, as is the case in linear fusion. While strictly controlling the computational cost of evaluation and achieving low-cost and rapid evaluation, it effectively improves the prediction accuracy of model performance and enhances the correlation and reliability between prediction results and actual performance. Through this fusion mechanism, this invention can accurately quantify the comprehensive performance of candidate architectures in daylily harvesting scenarios, taking into account segmentation stability under dense occlusion, the ability to distinguish subtle features of neck leaves, generalization performance in complex field environments, and the lightweight deployment requirements of edge harvesting equipment. It provides accurate comprehensive evaluation guidance for multi-objective architecture search, effectively shortens the NAS optimization iteration cycle, and provides core technical support for the engineering implementation of intelligent daylily harvesting.

[0013] Seventh, the MW-GCA daylily picking point positioning algorithm. In high-density daylily cultivation, the complex plant posture and terminal morphology, coupled with leaf occlusion, background interference, stem bending, and varied flower shapes and colors, make it difficult for traditional picking point localization algorithms to guarantee robustness. To address these issues, this algorithm integrates a corner point extraction strategy with an innovatively designed Mini-Window mechanism, aiming to achieve picking information acquisition in complex environments. Specifically, it extracts and merges the masked images of mature and immature regions. to obtain The geometric coordinates of the image corners are shown in the following mathematical expression. To address the issue of decreased positioning accuracy due to insufficient integrity of the edge region mask, the boundary corners were optimized. Subsequently, a corner detection strategy was used to locate potential picking points, and a corner filtering strategy was implemented for each connected region, focusing on eliminating invalid corners at the top of the plant and concentrating the analysis on the feature-rich areas at the bottom of the plant. An innovative Mini-Window mechanism was used to construct a local feature analysis window, which effectively shields the window from irrelevant background interference, allowing the algorithm to focus on the key features of the target region (algorithm flow is as follows). Figure 6 (as shown) ; Specifically, a Mini-Window refers to a tiny local window created near a corner point to analyze features in the vicinity and filter out corner points that meet the characteristics of a picking point. In this algorithm, a 60-pixel square is selected as the side length of the Mini-Window. After in-depth research on the corner points at the picking point, it was found that corner points meeting the requirements of a picking point share two common characteristics: First, if the Mini-Window is separated from the vertical axis of symmetry, the white area is mostly concentrated in the upper half, while the lower half is a black area; second, the upper half of the Mini-Window exhibits a "U" shape, which has three possible orientations: tilted to the left, tilted to the right, and centered. Based on the above neighborhood feature constraints, the precise coordinates of the picking point of the daylily plant are finally determined by filtering the local windows of candidate corner points based on geometric and grayscale distribution features.

[0014] Eighth, an algorithm for determining the daylily cutting angle based on Mini-Window. In the harvesting of daylilies, the choice of cutting angle has a crucial impact on the harvesting quality. A reasonable cutting angle can not only effectively reduce damage to the plants but also ensure that the harvested daylilies are of good appearance, thereby improving the overall economic benefits after the harvesting is automated. To address this issue, this invention conducts an in-depth analysis of the Mini-Window at the target corner point and constructs a cutting angle calculation method based on the Mini-Window. Specifically, firstly, by using the Mini-Window at the picking point, extract the two intersection points where the white area intersects with the Mini-Window boundary. , ,remember for and The midpoint, such as Figure 7 As shown. Then, connect them with a straight line. and This line segment represents the posture of the daylily plant corresponding to this picking point, and is denoted as... Finally, draw a line perpendicular to... The line segment is denoted as Through calculation The angle between the line segment perpendicular to the right and the line segment perpendicular to the right determines the final cutting angle at the picking point. The specific mathematical expression is as follows: ; ; ; ; ; ; in, and The function is used to obtain the left and right intersection points of the white area and the boundary of the Mini_Window. Functions used for drawing connections and A straight line, The function is used to draw the perpendicular line segment of a given line segment. The function is used to calculate the angle between a line segment and a horizontal line segment to the right.

[0015] Ninth, adapting the four-dimensional harvesting information of daylily picking to the conversion method of robotic arm control commands. To address the problem that existing detection methods in intelligent daylily harvesting scenarios cannot convert harvesting point information into 3D spatial information in the depth camera coordinate system, hindering the engineering implementation of the algorithms, this invention proposes a four-dimensional coordinate transformation method. This method combines a 3D spatial coordinate transformation algorithm with coordinate unification and motion calculation of the 4D harvesting information output by the MW-GCA algorithm. This generates 3D spatial coordinates and cutting angles that can be directly applied to robotic arm operations, enabling automated and precise harvesting of daylilies. This effectively reduces plant damage rates, ensures the integrity of harvested produce, promotes the application of intelligent harvesting technology, and enhances the economic benefits of the industry. Specifically, the four-dimensional harvesting information output by the MW-GCA algorithm covers three core data categories: two-dimensional pixel coordinates of the daylily harvesting point, three-dimensional depth information, and cropping angle. First, this invention converts the two-dimensional pixel coordinate information into three-dimensional spatial coordinates in the depth camera coordinate system. To achieve accurate mapping from pixel space to physical space, this invention collects the depth information of the daylily harvesting point using a depth camera. Based on the pinhole imaging model, the two-dimensional pixel coordinates Convert to 3D coordinates in depth camera coordinate system The specific formula is defined as follows: ; In the formula, and Depth camera in x direction and y Equivalent focal length in direction, This represents the optical center position in the pixel coordinate system. Through this coordinate transformation, this invention fuses the two-dimensional pixel position output by the MW-GCA algorithm with depth information, ultimately obtaining the complete three-dimensional spatial coordinates of the daylily picking point in the depth camera coordinate system. This provides precise spatial location data for planning the motion trajectory of the robotic arm, adapting to the precise positioning needs in the daylily picking scenario; To convert the cutting angle information into control commands that can be directly executed by the robotic arm's end effector, this invention employs a uniform linear mapping method to achieve precise conversion of angle values ​​into robotic arm servo pulse signals. This invention uses 0° to 180° as the effective input range for the cutting angle, corresponding to the output range of the robotic arm servo's pulse signals. The boundary pulse values ​​corresponding to 0° and 180° can be manually calibrated based on the actual robotic arm hardware parameters to ensure a high degree of compatibility between the mapping relationship and hardware control requirements. The specific mapping process is as follows: First, the input cutting angle is truncated, forcibly constraining angle values ​​exceeding the 0° to 180° range to within the range, preventing abnormal angles from causing the robotic arm to malfunction. Then, linear mapping is used to convert the angle to a pulse value, achieving a one-to-one correspondence between angle and pulse. Finally, the mapped pulse value is rounded to generate an integer pulse signal that meets the servo control requirements, which is directly used to drive the robotic arm's end effector to complete the precise cutting of daylilies. This linear mapping method can flexibly adjust the boundary pulse value according to the hardware parameters of different robotic arms, adapt to the control requirements of different models of harvesting equipment, effectively improve the accuracy of cutting angle control, avoid plant damage and appearance destruction caused by angle deviation, and ensure the integrity and quality of daylily harvesting. The four-dimensional coordinate transformation method proposed in this invention, through the synergistic effect of three-dimensional spatial coordinate transformation and angle-pulse linear mapping, completely transforms the four-dimensional harvesting information output by the MW-GCA algorithm into control commands that can be directly executed by the robotic arm. This opens up the technical link from vision detection to robotic arm operation, solves the core problem that existing algorithms are difficult to implement in engineering, realizes automated and precise harvesting of daylilies, effectively reduces plant damage rate, ensures the integrity of harvested products, provides key technical support for the engineering application of intelligent daylily harvesting, and significantly improves the economic benefits of the industry.

[0016] Tenth, a real-time coordinate-attitude mapping analytical algorithm for daylily harvesting robotic arms. To address the issues of insufficient positioning and attitude control accuracy in intelligent daylily harvesting, where the 3D harvesting coordinates from depth cameras cannot be accurately mapped to the robotic claw coordinate system, and the harvesting angle is difficult to convert into rotation control signals, this invention proposes a kinematic analysis algorithm for robotic arms adapted to this scenario. This algorithm, through kinematic modeling and analysis, maps the 3D harvesting coordinates from the depth camera to the robotic claw coordinate system. Simultaneously, it converts the trimming angle information into rotation pulse control values ​​for the robotic claw, enabling precise scheduling and attitude control of the robotic arm. This effectively improves harvesting positioning and attitude accuracy, reduces plant damage, ensures the quality of the finished product, and provides reliable motion control support for the successful implementation of intelligent daylily harvesting. Specifically, this invention addresses the motion control requirements of a six-DOF robotic arm in daylily harvesting scenarios, constructing a complete kinematic analysis chain from vision detection to joint control. The specific process is as follows: First, the robotic arm is fixed at a preset picking point to detect the pose. Image information of the daylily picking point is acquired using a depth camera. The MW-GCA picking point detection algorithm and four-dimensional coordinate transformation method described in this invention are then used to obtain the three-dimensional spatial coordinates and clipping angle information in the depth camera coordinate system. Subsequently, based on the hand-eye calibration results of the depth camera and the robotic arm, the three-dimensional picking point coordinates in the depth camera coordinate system are converted to the end effector target pose in the robotic arm's world coordinate system using a fixed coordinate transformation relationship between the depth camera coordinate system and the robotic arm's world coordinate system. This provides accurate target input for subsequent inverse kinematics solving, meeting the precise positioning requirements of the complex environment in daylily fields. To address the problems of traditional inverse kinematics solutions relying on complex matrix operations, high computational cost, and difficulty in meeting the real-time requirements of harvesting, this invention employs geometric analysis to perform inverse kinematics analysis on a robotic arm. By omitting redundant calculations of the bottom gimbal rotation joints, the three-dimensional spatial motion problem is simplified into a two-dimensional planar kinematic analysis, significantly reducing computational complexity, improving solution efficiency, and adapting to the real-time operational requirements of intelligent daylily harvesting. The core objective of this inverse kinematics analysis is: given the spatial position and orientation of the robotic arm's end effector, to solve for the rotational parameters of each joint, providing precise joint drive commands for the robotic arm's motion control. Specifically, based on the inherent mechanical structural properties of the robotic arm (length of each link) , , Given the known conditions (e.g., the target coordinates), we first simplify the target coordinates and define intermediate variables. m , n The specific formula is defined as follows: ; ; In the formula, x , y The coordinates of the target at the end of the robotic arm. This refers to the attitude angle of the end effector. m , n Substituting into the kinematic constraint equations, we can further simplify to obtain the link length constraint formula, which is defined as follows: ; By solving this quadratic equation, the joint angle can be obtained. The sine value is defined by the following formula: ; In the formula, a , b , c The coefficients are calculated based on intermediate variables and link lengths, and the specific formulas are defined as follows: ; ; ; Based on the above results, this invention further completes the full analysis of the rotation angles of each joint of the robotic arm. For the five joints excluding the opening and closing of the robotic gripper, the rotation angle parameters of each joint are solved sequentially, and the specific formulas are defined as follows: ; In the formula, , , The three-dimensional coordinates of the robotic arm's end effector in the world coordinate system. , For the DH parameters of the robotic arm, These are the link attitude parameters. , The parameters related to the end effector attitude are shown in Table 4. Using the above geometric analytical method, the precise rotation angle parameters of each joint of the robotic arm can be quickly obtained, replacing traditional complex matrix operations, significantly improving solution efficiency and accuracy, and avoiding the problems of computational redundancy and poor real-time performance of traditional methods. Table 4 DH Parameters of the Robotic Arm ai-1 αi-1 di θi (θmin, θmax) 1 0 0 0 θ1 θ1(-120,120) 2 -π / 2 0 0 θ2 θ2(-180,0) 3 0 0.13 0 θ3 θ3(-120,120) 4 0 0.13 0 θ4 θ4(-200,20) 5 -π / 2 0 0 θ5 θ5(-120,120) The proposed analytical algorithm for robotic arm kinematics simplifies the inverse kinematics solution process through geometric analysis, achieving precise mapping from the three-dimensional coordinates of the depth camera to the joint control of the robotic arm. Simultaneously, it effectively converts the cutting angle into control pulses, effectively solving the problem of insufficient positioning and posture control accuracy in intelligent daylily harvesting. This significantly improves the real-time performance and accuracy of robotic arm motion control, reduces plant damage rate, ensures the integrity of harvested daylilies, provides core motion control technology support for the engineering implementation of intelligent daylily harvesting, and promotes the development of the intelligent daylily harvesting industry.

[0017] This invention constructs the TYUST-Daylily semantic segmentation dataset for daylilies in a dense field environment. Pixel-level annotations are applied to 700 images using LabelMe, classifying daylilies into three categories based on their growth stage: mature, immature, and overmature. Data augmentation is performed through random brightness adjustment, noise injection, and horizontal / vertical flipping, increasing the dataset to 1400 images. Finally, the training and validation sets are non-overlapping and divided in an 8:2 ratio, providing standardized data support for subsequent experiments.

[0018] To verify the prediction accuracy and relevance of the low-cost evaluation strategy proposed in this invention, this embodiment relies on the NAS-Bench-201 benchmark search space. 1000 different network architectures are randomly sampled from the benchmark search spaces corresponding to three typical datasets: CIFAR-10, CIFAR-100, and ImageNet16-120, as evaluation samples, covering different complexities and structural types to ensure sample diversity. The sampled architectures are evaluated using the low-cost neural architecture performance evaluation strategy based on multi-dimensional collaboration and ranking-driven proposed in this invention to obtain prediction scores. Simultaneously, the actual test accuracy of each architecture on the corresponding dataset is collected as a benchmark. The two types of data are visualized using a scatter plot. The results are as follows: Figure 8 As shown in the visualization results, it can be intuitively observed that the predicted scores and the actual accuracy of the architectures output by this evaluation strategy exhibit a highly positive correlation distribution across all three datasets, with no significant deviation trend. Quantitative analysis further indicates that the Kendall correlation coefficient between the predicted scores and the actual accuracy is 0.7120 and the Spearman correlation coefficient is 0.8912 for the CIFAR-100 dataset; 0.7052 and 0.8855 for the CIFAR-100 dataset; and 0.6981 and 0.8663 for the ImageNet16-120 dataset. This fully demonstrates that the low-cost evaluation strategy proposed in this invention can accurately characterize the comprehensive performance of neural architectures, effectively achieve candidate architecture performance evaluation while controlling computational overhead, improve the efficiency and reliability of architecture screening, and provide core technical support for the efficient screening of high-quality architectures in neural architecture search.

[0019] To further verify the technical advancement of the low-cost evaluation strategy proposed in this invention, this embodiment conducts a cross-dataset comparative analysis between the strategy and existing advanced methods based on an experimental sample size of 1000. The specific comparison results are shown in Table 5. Experimental results show that the low-cost neural architecture performance evaluation strategy based on multidimensional collaboration and ranking-driven proposed in this invention significantly outperforms the comparative methods in terms of Kendall correlation coefficient τ and Spearman correlation coefficient ρ on three typical datasets: CIFAR-10, CIFAR-100, and ImageNet16-120. Specifically, on the CIFAR-10 dataset, the proposed strategy achieves a τ value of 0.7120, with an absolute difference of 0.1193 compared to the second-best performing NASHOT method, representing a relative improvement of approximately 20.1%. The ρ value is 0.8912, 14.5 percentage points higher than the NASHOT method. On the more complex ImageNet16-120 dataset, the proposed strategy maintains a high τ value of 0.6981, 0.0864 higher than the NASHOT method. These cross-dataset comparisons fully demonstrate that the proposed method achieves superior correlation metrics on all test datasets, showcasing the outstanding advantages and advancements of this nonlinear evaluation strategy in neural architecture performance characterization and evaluation. It enables accurate characterization and efficient screening of candidate architecture performance while controlling extremely low computational overhead, providing core technical support for rapid optimization in neural architecture search. Table 5 Comparison of different low-cost assessment methods

[0020] To systematically analyze the sensitivity of the low-cost evaluation strategy proposed in this invention to changes in sampling scale, this embodiment relies on the NAS-Bench-201 benchmark search space and conducts comparative experiments on three datasets: CIFAR-10, CIFAR-100, and ImageNet16-120, selecting six different sampling scales of 10, 100, 500, 1000, 2000, and 3000 respectively. The Kendall correlation coefficient τ of the strategy of this invention and existing advanced evaluation methods are compared and analyzed at each sampling scale. The comparison results of each method are summarized in the form of a radar chart in Figure 9. In this radar chart, each axis corresponds to a different sampling scale, and the scale on the axis represents the Kendall correlation coefficient value at the corresponding sampling scale. Experimental results show that, under different sampling scales of the CIFAR-10, CIFAR-100, and ImageNet16-120 datasets, the radar chart area corresponding to the strategy proposed in this invention is significantly larger than that of the comparative methods, intuitively demonstrating the superior relevance performance of this strategy at various sampling scales. Further analysis shows that, across the entire range from a minimum of 10 to 3000, the prediction score output by the strategy of this invention consistently maintains a relatively high and stable relevance level without significant fluctuations or sudden drops. In contrast, the relevance indicators of some comparative methods show significant fluctuations with changes in sampling scale, or the overall relevance coverage is relatively limited, making it difficult to adapt to the evaluation needs of different sampling scales. The above results fully demonstrate that the low-cost neural architecture performance evaluation strategy based on multi-dimensional collaboration and ranking-driven proposed in this invention possesses strong stability and robustness when the number of samples changes. This characteristic makes it highly practical in real-world neural architecture search applications where computational resources are limited and the sample size is uncertain. It can adapt to the architecture evaluation needs of different scales and effectively ensure the efficiency and reliability of architecture selection in resource-constrained scenarios.

[0021] To intuitively verify the actual contribution of each proxy metric to model performance evaluation in the low-cost neural architecture performance evaluation strategy based on multidimensional collaboration and ranking driven proposed in this invention, this embodiment conducts a systematic ablation experiment under the condition of uniformly sampling 1000 network architectures. The experimental results are shown in Table 6. Each column in the table corresponds to whether the corresponding proxy metric is included, the metric aggregation type, and the Kendall correlation coefficient τ under the three datasets CIFAR-10, CIFAR-100, and ImageNet16-120. Experimental results show that when using only a single surrogate metric, the Kendall correlation coefficient τ for each dataset is in a low range, fully demonstrating that a single-dimensional metric cannot comprehensively and accurately characterize the overall performance of the neural architecture. As the number of surrogate metrics gradually increases, the τ values ​​for the three datasets show a steady upward trend, verifying that the four types of surrogate metrics designed in this invention—model perturbation robustness, feature space isotropy, parameter-gradient joint uncertainty, and gradient-layer structure regularity—have complementary advantages and can synergistically enhance the architecture evaluation performance. Notably, after using the ranking-based nonlinear multi-metric geometric fusion mechanism proposed in this invention, the τ values ​​for all three datasets reach the experimental optimum level, with 0.7120 for the CIFAR-10 dataset, 0.7052 for the CIFAR-100 dataset, and 0.6981 for the ImageNet16-120 dataset. This result is significantly better than the linear fusion method using the same surrogate metrics. For example, in the CIFAR-10 dataset, the τ value of the nonlinear fusion method in this invention is about 0.0137 higher than that of the linear fusion method. In summary, the four types of proxy indicators constructed in this invention can provide complementary and diverse architectural evaluation information. Compared with the traditional linear aggregation method, the proposed nonlinear geometric fusion method can integrate the effective information of each proxy indicator more efficiently, significantly improving the accuracy of the model performance prediction score. The complementary design of the indicator system and the reasonable selection of the fusion method together constitute the core technical foundation for the low-cost evaluation strategy of this invention to achieve high predictive relevance. Table 6 Low-cost evaluation indicators for ablation experiments Index Srobust Sisotropy Suncertainty Suniformity FM CIFAR-10 CIFAR-100 ImageNet16-120 1 ✔ - 0.4897 0.5029 0.3451 2 ✔ - 0.4192 0.3683 0.3993 3 ✔ - 0.5553 0.5798 0.5485 4 ✔ - 0.5776 0.6055 0.5842 5 ✔ ✔ NL 0.5996 0.5695 0.5477 6 ✔ ✔ NL 0.6088 0.6259 0.5466 7 ✔ ✔ NL 0.6178 0.6368 0.5560 8 ✔ ✔ NL 0.5907 0.6187 0.5884 9 ✔ ✔ ✔ NL 0.6870 0.6715 0.6597 10 ✔ ✔ ✔ NL 0.6826 0.6759 0.6749 11 ✔ ✔ ✔ NL 0.6303 0.6521 0.5965 12 ✔ ✔ ✔ ✔ L 0.6983 0.6701 0.5919 13 ✔ ✔ ✔ ✔ NL 0.7120 0.7052 0.6981

[0022] To further verify the effectiveness of the proposed evaluation strategy in practical neural architecture search, this embodiment uses NAS-Bench-201 as the benchmark search space. Under a fixed sampling size of 1000, comparative experiments are conducted on the CIFAR-10, CIFAR-100, and ImageNet16-120 datasets, comparing the proposed method with existing training-free evaluation methods. Each evaluation method is run independently three times, and the performance of the best model obtained in each round of experiment is used as the search result. The experimental results are presented in the form of mean ± standard deviation, and the accuracy (Optimal) of the best model in the candidate model set is used as the experimental reference standard. The final experimental results are shown in Table 7. The experimental results show that the proposed strategy achieves optimal or near-optimal performance on all datasets, fully demonstrating robust high-quality architecture selection capabilities. On the CIFAR-10 dataset, the proposed strategy achieved an accuracy of 93.41 ± 0.31, outperforming the second-best performing ZiCo method by 0.19 percentage points, and only 0.79 percentage points behind the theoretical optimum—a significantly smaller gap than other comparative methods. On the more challenging ImageNet16-120 dataset, the proposed strategy maintained its leading performance with an accuracy of 44.33 ± 1.62, outperforming ZiCo and NASWOT by 0.30 and 0.21 percentage points respectively. On the CIFAR-100 dataset, the proposed strategy achieved an accuracy of 70.23 ± 0.41, placing it in the high-performance gradient range alongside other training-free evaluation methods such as SynFlow. Crucially, the proposed strategy not only stably selects high-performance network architectures across various datasets but also ensures that the search results maintain a high degree of consistency with the theoretical optimum. The results fully demonstrate that the multidimensional evaluation system designed in this invention can effectively capture the core features that determine the final performance of the model, and achieve near-optimal architecture search results under almost no training conditions, providing a solution that balances search efficiency and reliability for automated model architecture design under various resource-constrained scenarios. Table 7 Performance Comparison of Different Low-Cost Evaluation Strategies .

[0023] To fully verify the technical effectiveness and engineering practicality of the proposed H²D²-MONAS daylily segmentation model automatic architecture method in actual daylily harvesting scenarios, this embodiment conducts a full-process verification experiment based on the self-constructed TYUST-Daylily daylily segmentation dataset. This embodiment deploys the H²D²-MONAS framework in the aforementioned standard experimental environment. Using the TYUST-Daylily dataset as input, it leverages a hierarchical heterogeneous multi-source fusion search space for daylily segmentation, a multi-objective neural architecture search strategy guided by dual-domain semantic collaboration, and a low-cost neural architecture performance evaluation strategy based on multi-dimensional collaboration and ranking-driven optimization. End-to-end automatic optimization of the optimal semantic segmentation network architecture is conducted, ultimately yielding the optimal semantic segmentation model suitable for complex scenarios such as dense occlusion and high similarity of neck leaf features in daylilies. This optimal model has only 7.839M parameters and 4.783G of floating-point operations (FLOPs). On the TYUST-Daylily test set, it achieves a mean intersection-union ratio (mIoU) of 62.55%, a mean pixel accuracy (mPA) of 75.39%, and a precision of 95.35%. While meeting the lightweight deployment requirements of agricultural intelligent equipment, it also ensures high precision in daylily segmentation, fully verifying the efficiency and practicality of the NAS framework of this invention. The 63-dimensional encoding sequence corresponding to the optimal architecture is [2, 1, 7, 1, 1, 3, 2, 4, 5, 8, 1, 18, 9, 9, 8, 5, 2, 1, 3, 1, 3, 2, 7, 4, 0, 7, 5, 2, 3, 5, 6, 7, 2, 2, 0, 1, 2, 2, 6, 2, 5, 4, 6, 19, 5, 5, 6, 10, 3, 10, 8, 7, 9, 2, 7, 4, 4, 6, 10, 6, 4, 5, 8]. This sequence fully represents the hierarchical heterogeneous cell structure, node topology connections, and operator configuration of the architecture, providing a precise structural basis for model reproduction, deployment, and secondary optimization. Based on the high-precision semantic segmentation mask image output by this optimal architecture, this embodiment further utilizes the MW-GCA intelligent localization algorithm for daylily picking points proposed in this invention. Through Mini-Window local window traversal filtering, VASU-based "U"-shaped feature window matching, and attitude line fitting, it accurately extracts the four-dimensional information of the picking point pixel coordinates and cutting angle of the daylily plant to be cut. In actual robotic arm testing, this solution demonstrated excellent real-time performance: the processing speed for daylily segmentation was approximately 12 frames per second (FPS), and the single-frame picking point localization speed was approximately 50 milliseconds. Even in embedded device deployment scenarios, it can meet the real-time requirements of automated picking. Subsequently, this four-dimensional picking information is transformed into executable motion commands for the robotic arm's end effector through a three-dimensional spatial coordinate transformation algorithm and the robotic arm kinematics analysis module. This drives the 6-DOF robotic arm to move the flexible gripper to complete the automated and precise picking of the daylily. The actual simulated picking effect is as follows: Figure 10 As shown. This end-to-end solution can effectively avoid problems such as delayed harvesting and damage to the appearance caused by the time-sensitive nature, physical strength and experience limitations of manual harvesting, significantly reducing economic losses in agricultural production, and providing core technical support for the intelligent transformation and upgrading of the entire daylily industry. All experiments in this invention were conducted under standard experimental conditions. The specific hardware and software configurations are as follows: For general computing hardware, the processor was a 14th Gen Intel® Core™ i5-14600KF with a clock speed of 3.50GHz; the graphics card was an NVIDIA GeForce RTX 4070 Ti SUPER; the memory configuration was 32GB; the robotic arm control hardware was a Jetson Orin NX (8GB); the visual acquisition hardware was an Orbbec Gemini Plus depth camera; and the software development environment was PyTorch 3.8. This experimental environment provided stable and reliable support for the technical verification and performance testing of the invention, ensuring the authenticity and reproducibility of the experimental results.

[0024] This invention possesses strong practicality and wide applicability. Its core application is to semantic segmentation and automated harvesting of daylilies, specifically addressing the technical pain points of existing daylily harvesting processes, such as high reliance on manual labor, low efficiency, and significant losses. In practical industrial applications, this invention effectively avoids problems such as delayed harvesting and damage to the appearance of daylilies caused by limitations imposed by experience, physical condition, and working time during manual operation, significantly reducing economic losses in agricultural production. Simultaneously, by replacing traditional manual harvesting with automation technology, it powerfully promotes the engineering implementation and industrialization of automated crop harvesting technology, improving the production efficiency and intelligent level of the daylily industry. From the perspective of macro-industry and national strategy, the implementation and promotion of this invention actively aligns with the national strategic orientation of promoting the healthy development of the intelligent manufacturing industry and encouraging technological and independent innovation. Based on the needs of agricultural intelligent transformation and upgrading, it empowers key links in agricultural production through core technologies, injects core impetus into the transformation of the agricultural industry towards intelligence and modernization, helps to build an efficient, green, and intelligent modern agricultural production system, and promotes the high-quality and sustainable development of characteristic economic crops such as daylilies. It should be noted that the above content is merely a specific embodiment of the present invention, intended to help those skilled in the art understand and implement the present invention, and is not intended to limit the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims. Any reasonable improvements, equivalent substitutions, modifications, and optimizations made based on the technical concept and core principles of the present invention, as long as they do not depart from the essence of the core technical solution of the present invention, should be included in the scope of protection of the present invention. In addition, adaptive adjustments and conventional modifications made by those skilled in the art without violating the principles of the present invention or departing from the technical concept of the present invention are also within the scope of protection of the present invention.

Claims

1. An adaptive daylily segmentation and four-dimensional picking point localization method based on heterogeneous dual-domain collaboration, characterized in that, Includes the following steps: Step 1: Construct an automatic architecture method for the H²D²-MONAS daylily segmentation model. This method is based on the encoder-decoder full-link heterogeneous staged search unit for daylily segmentation tasks. It completes the architecture representation through a high-degree-of-freedom topological coding strategy that supports single-node multi-precursor node connections. It constructs a multi-objective neural architecture search strategy guided by dual-domain semantic collaboration through an adaptive fusion congestion calculation mechanism of the target performance domain and the architecture structure domain, a dual-mode crossover operator adapted to cell-level segmentation coding, and a topology-operation cooperative mutation operator. Combined with a multi-dimensional collaborative model performance evaluation strategy and a ranking-based nonlinear geometric fusion mechanism, it completes the evaluation and screening of candidate architectures and autonomously searches for the optimal semantic segmentation model adapted to the daylily picking scenario. Step 2: Input the daylily mask output by the optimal semantic segmentation model into the MW-GCA daylily picking point localization algorithm. Through corner detection and Mini-Window local window filtering, obtain the two-dimensional pixel picking point coordinates of the daylily to be picked. Step 3: The algorithm for determining the daylily cutting angle based on Mini-Window analyzes the geometric features of the local window at the picking point and calculates the cutting angle of the corresponding picking point. Step 4: Convert the two-dimensional pixel coordinates and depth information of the picking point into three-dimensional spatial coordinates in the depth camera coordinate system using a four-dimensional coordinate transformation method. Combine this with the kinematic analysis algorithm of the robotic arm to complete the coordinate mapping and joint pulse solution, thereby driving the robotic arm to complete automated and precise picking.

2. The adaptive daylily segmentation and four-dimensional picking point localization method based on heterogeneous dual-domain collaboration according to claim 1, characterized in that, The encoder-decoder end-to-end heterogeneous phased search unit mentioned in step one is specifically: The encoder is equipped with a three-level feature extraction unit. Each level unit consists of a cascaded ResCell, a dual-depth feature extraction FECell, and an attention AttCell. The ResCell contains two computing nodes to realize feature downsampling and basic feature preservation. The FECell realizes multi-branch feature collaborative extraction through a five-node multi-preorder connection structure. The AttCell generates attention weights through four-node parallel operation to highlight the key features of daylily. The decoder uses two-stage FPACell heterogeneous decoding units: FPACell S1 performs primary multi-scale feature fusion, and FPACell S2 completes deep multi-scale feature fusion, which completes the features of the occluded region and enhances the distinguishability of the neck lobe features, forming a heterogeneous staged search space for the encoder-decoder whole link.

3. The adaptive daylily segmentation and four-dimensional picking point localization method based on heterogeneous dual-domain collaboration according to claim 1, characterized in that, The high-degree-of-freedom topology coding strategy that supports single-node multi-precursor node connections described in step one is as follows: The complete segmentation architecture is represented by a 63-bit integer encoding, where the three-level feature extraction stage corresponds to a 48-bit encoding. Each 16-bit stage consists of a 2-bit operation encoding for ResCell, a 10-bit operation encoding for FECell, and a 4-bit operation encoding for AttCell. In the decoding stage, FPACell S1 corresponds to a 6-bit encoding and FPACell S2 corresponds to a 3-bit encoding. FECell's encoding uses a combination of "operation code + connection code". Odd-numbered bits are operation codes, corresponding to the candidate operations of each node, while even-numbered bits are connection codes, representing the connection relationship between each node and its predecessor nodes. The decimal value of the connection code corresponds to the binary form, and the number of binary bits is consistent with the node's sequence number. Taking node x3 as an example, its connection code corresponds to a 3-bit binary number, which, from left to right, represents the connection status with predecessor nodes 0, 1, and 2 (1 indicates connection, 0 indicates no connection). If the connection code is 7, the binary form is 111, meaning that node x3 is connected to predecessor nodes 0, 1, and 2. FECell's connection encoding represents node connections by converting decimal to binary, with the node number matching the number of binary bits. A single node can connect to multiple preceding nodes simultaneously, breaking through the traditional single-node, single-precedence connection constraint and enabling topological representations such as residual jump connections, multi-branch parallelism, and cross-level multi-scale fusion.

4. The adaptive daylily segmentation and four-dimensional picking point localization method based on heterogeneous dual-domain collaboration according to claim 1, characterized in that, The adaptive fusion congestion calculation mechanism between the target performance domain and the architecture domain mentioned in step one is as follows: The congestion level is obtained by adaptive weighted fusion, and the fusion formula is as follows: ； In the formula, For the final integration of individual crowding, For the target domain congestion, For the congestion of the structural domain, The weights are dynamically adaptive, and the crowding levels of both domains are normalized to the [0,1] interval. The larger the value, the more crowded the individual distribution under the corresponding dimension. Individuals with lower crowding levels are selected first during the screening process. Target domain crowding is used to measure the density of candidate architectures in the optimization target space. The optimization target includes two metrics: the number of model parameters and the low-cost evaluation score. The calculation first performs min-max normalization on the target values ​​of individuals within the same non-dominated front to eliminate dimensional differences. Then, for each individual, the two nearest neighbors in the target space with the closest Euclidean distance are selected, and the original target domain crowding is calculated based on the nearest neighbor distance, as shown in the following formula: ； In the formula, For individuals i and neighbors j In the Euclidean distance of the target space, To prevent the use of tiny constants with a denominator of zero, for individuals located at the boundary of the target space, their original crowding degree is directly set to 0. Then, the original target domain crowding degree of all individuals is normalized using a min-max method to obtain the normalized target domain crowding degree. ; Domain crowding is used to measure the similarity of candidate architectures in hierarchical heterogeneous cell topologies. This invention uses the 63-dimensional encoding vectors corresponding to each architecture as the basis for structural representation to achieve a digital description of the topology. Subsequently, the cosine similarity between the encoding vectors of different individuals is calculated. For each individual, the two nearest neighbors with the highest similarity are selected, and the original domain crowding is calculated based on structural similarity. The calculation formula is as follows: ； In the formula, For individuals i and neighbors j The structural cosine similarity is calculated. After the original crowding degree is calculated, the results of individuals within the same front are normalized using the min-max method to obtain the normalized domain crowding degree. ; To achieve dynamic adaptive allocation of weights, this invention calculates the fusion weights based on the distribution variance of the two-domain crowding degree within the current non-dominated frontier. The calculation formula is as follows: ； In the formula, and These represent the variances of the target domain crowding and the structural domain crowding within the current frontier, respectively. These weights can be automatically adjusted based on population distribution. When the structural domain distribution is more dispersed, the target domain weight is increased to ensure diversity in segmentation performance; when the target domain distribution is more dispersed, the structural domain weight is strengthened to maintain architectural topological diversity and avoid excessive homogeneity of solution sets.

5. The adaptive daylily segmentation and four-dimensional picking point localization method based on heterogeneous dual-domain collaboration according to claim 1, characterized in that, The dual-mode crossover operator and topology-operation cooperative mutation operator adapted to cell-level segmented encoding mentioned in step one are specifically as follows: The dual-mode crossover operator uses a total crossover probability of 0.9 and fuses cell-level crossover and uniform crossover in a ratio of 0.6:0.

4. Cell-level crossover randomly exchanges complete cell coding segments, while uniform crossover performs fine-grained exchanges on single gene loci. After crossover, a validity repair is performed to lock the core coding position. The topology-operation cooperative mutation operator binds the topology connection bits and operation bits to mutation. The mutable connection bits perform binary bit flipping to achieve topology perturbation, and the operation bits are randomly updated within the candidate operation set. The mutation process locks the rigid constraint bits to ensure the integrity of the cell structure and the encoded semantics.

6. The adaptive daylily segmentation and four-dimensional picking point localization method based on heterogeneous dual-domain collaboration according to claim 1, characterized in that, The performance evaluation strategy for the multidimensional collaborative model described in step one is as follows: Regarding the robustness of the model to perturbations, firstly, addressing the common problems in daylily harvesting scenarios such as image information loss and partial occlusion, this invention perturbs the original input image through a center cropping operation to generate perturbed test input samples. Its generation formula is defined as follows: ； in, Center-based clipping function; Image scaling function; input tensor This represents a single batch of input images, where dimensions B, C, H, and W represent the batch size, number of channels, image height, and width, respectively. During the evaluation process, this invention uses batches containing 128 samples as input to ensure that the evaluation results fully characterize the overall distribution features of the dataset. Subsequently, to eliminate the dimensional differences between different feature dimensions and improve sensitivity to subtle changes in feature distribution, this invention uses the rank correlation coefficient principle to measure distribution consistency. Specifically, the feature vectors extracted from the original input sample and the perturbed sample by the model are sorted by element value and assigned corresponding rank coefficients, thereby constructing the rank matrix corresponding to the original input. Rank matrix corresponding to the perturbation input Based on this, the present invention calculates the average rank correlation coefficient of a single batch of samples. To quantify the distributional consistency between two-rank matrices, the calculation formula is as follows: ； In the formula, and Let be the mean of the rank vector of the b-th sample. To prevent the smoothing constant from having a denominator of zero; Finally, to map the evaluation results to a unified quantization range, this invention introduces an exponential mapping function to transform the rank correlation coefficient into a robust evaluation score in the range of 0 to 100. The specific formula is defined as follows: ； For the model feature space to be isotropic, the single-batch input image sampling strategy consistent with the model perturbation robustness evaluation is used, and the input tensor... Each training batch consists of 128 images randomly sampled from the dataset, ensuring that the evaluation results fully reflect the overall distribution characteristics of the dataset. It relies on model forward propagation and hook registration to collect images from the network's front... L The output feature tensor of the first main block; for the first... l Each main block has output characteristics ,in For the number of channels, and Represent the height and width of the feature map, and reshape the feature tensor into a feature matrix. ,in The total number of feature samples; Next, the characteristic matrix The centered feature matrix is ​​obtained by performing a centralization process. Then add regularization parameters. Calculate the regularized covariance matrix to avoid matrix singularity issues caused by insufficient samples. Based on this, calculate the real symmetric covariance matrix. Perform eigenvalue decomposition to obtain eigenvalues. Then set a threshold. By performing nonnegative adjustment and L1 normalization on the eigenvalues, the principal component probability distribution is obtained. The specific formula is defined as follows: ； Furthermore, the probability distribution of the principal components is used to calculate the first... l Feature representation entropy of each main block To quantify the dispersion of the main block's feature distribution, the specific formula is defined as follows: ； Finally, all L The feature representation entropy of each principal block is accumulated to obtain the overall feature space isotropic evaluation score of the architecture. The specific formula is defined as follows: ； For parameter-gradient joint uncertainty, the multi-factor weighting strategy proposed in this invention uses the two-dimensional matrix flattened from the parameters of each layer. W Based on the computational foundation, the comprehensive importance weight of each layer is quantified by integrating key weighting factors from four dimensions. Among them, the basic scale weight is defined as... , For parameter matrix W The total number of elements, which characterizes the computational complexity of the model through the total number of parameters, is a fundamental dimension for importance assessment; the network depth coefficient is defined as... ,in b The index of the main block where the layer is located. B The total number of main blocks is used to model the position of the layers in the gradient propagation path. The deeper the network layer, the more significant its contribution to the model output, adapting to the deep network requirements for daylily feature extraction; the module stacking coefficient is defined as... , The number of normal unit stackings is used to characterize the feature accumulation effect when modules are stacked. The more stackings, the higher the weight, adapting to the stacking structure requirements of multi-scale feature extraction; functional role weight. The contribution of different layers to the network function is assigned values: the downsampling layer, responsible for preserving discriminative information and significantly impacting the final model's performance and efficiency, is assigned the highest weight of 1.60; the feature extraction layer, primarily tasked with extracting subtle features from the daylily's neck and leaves, is assigned a weight of 1.35, higher than the baseline; the preprocessing layer, used to ensure the stability of the input data, is assigned a weight of 1.10, slightly higher than the baseline; the classification layer, heavily reliant on preceding feature outputs, has a relatively low weight, set at 0.90; the global pooling layer and other layers serve only as bridges for information exchange, and are assigned the baseline weight of 1.

00. Finally, the four weight factors are weighted and fused using the following formula to obtain the comprehensive importance weight of each layer: ； For weight matrices of different scales, this invention first performs singular value decomposition on them to obtain singular value sequences. Then, a cumulative energy threshold was set. effective rank Defined as satisfying The smallest positive integer, where Let be the full rank of the matrix; if the matrix norm This indicates that the information contribution of this layer is weak during the initialization phase, and the effective rank ratio of this layer is set to be... Otherwise, the effective rank ratio is defined as Finally, this invention uses a comprehensive importance weight for each layer. I As a weighting factor, the rank ratio of the effective layer is used. A weighted average is then performed to obtain a score for the joint uncertainty of the model parameters and gradients. The specific mathematical expression is as follows: ； Regarding the gradient-layer structure regularity, specifically, to more stably obtain the statistical features of the model, this invention generates... N Group( N =5) Input samples conforming to a standard normal distribution and their corresponding target labels are used to generate data that is input into the model to be evaluated to perform backpropagation, thereby obtaining the gradients of the trainable parameters in the candidate architecture. ( L (The loss value of the loss function). For different parameters W and its gradient First, calculate its direction perception uncertainty. The specific formula is defined as follows: ； In the formula, Indicates the learning rate. It is a smoothing factor; Based on this, N Valid data obtained from the second sampling Merge into vectors (Remove items with absolute values ​​less than) (approximately zero value), calculate the mean of its absolute value sequence. and standard deviation .like Then, the absolute value of each element is obtained through a Gaussian distribution. Fit the probability distribution; if This indicates a highly concentrated distribution, and a uniform distribution is used for fitting. Subsequently, the fitted probability distribution is... Calculate the gradient entropy of this layer. The specific formula is defined as follows: ； Furthermore, to further quantify the differentiated impact of different layer types on the model training process and to adapt to the hierarchical structure characteristics of the daylily semantic segmentation network, this invention introduces a role multiplier. (Numerical and parameter-gradient joint uncertainty dimensions are consistent), depth scaling (in b This refers to the index of the main block where the layer is located. B (Total number of main blocks) and stacking ratio (in s (The number of stacked Normal blocks), for different layers in the model. We perform weighting to obtain the weighted hierarchical entropy value. The specific formula is defined as follows: ； Finally, for all effective layers ( The total weighted entropy is obtained by summing the weighted entropies of elements with length ≥ 10. And further calculate the average weighted entropy (in M (This represents the number of effective layers). If the value of M is 0, the score is returned directly. Otherwise, calculate the intermediate score S, and then normalize the intermediate score to the interval [0,100], which will be used as the final model gradient-layer structure regularity score. The formula for calculating the median score S is defined as follows: ； The four indicators comprehensively evaluate the candidate architecture from the dimensions of segmentation robustness, feature representation, generalization performance, and lightweight structure, adapting to the complex field environment and equipment deployment requirements of daylily harvesting scenarios.

7. The adaptive daylily segmentation and four-dimensional picking point localization method based on heterogeneous dual-domain collaboration according to claim 1, characterized in that, The ranking-based nonlinear geometric fusion mechanism described in step one is as follows: A score matrix is ​​constructed from the scores of the four evaluation indicators. The scores of each indicator are then sorted and normalized using a dense ranking method to obtain the rank ratio matrix. The four-dimensional geometric mean of the rank ratio vector of each individual architecture is calculated to obtain the final comprehensive evaluation score, as shown in the formula: ； By eliminating the defects of linear fusion and retaining the relative advantages and disadvantages of each indicator, we can achieve low-cost and high-precision performance prediction.

8. The adaptive daylily segmentation and four-dimensional picking point localization method based on heterogeneous dual-domain collaboration according to claim 1, characterized in that, The MW-GCA daylily picking point location algorithm described in step two is as follows: First, extract the masked merged image of mature and immature regions. to obtain The geometric coordinates of the image corners are expressed mathematically as follows. To address the issue of decreased positioning accuracy due to insufficient integrity of the edge region mask, the boundary corners were optimized. Subsequently, a self-developed corner detection strategy was used to locate potential picking points. A corner filtering strategy was implemented for each connected region, focusing on eliminating invalid corners at the top of the plant and concentrating the analysis on the feature-rich areas at the bottom of the plant. An innovative Mini-Window mechanism was used to construct a local feature analysis window, which effectively shields the window from irrelevant background interference, allowing the algorithm to focus on the key features of the target region. ; Specifically, a Mini-Window refers to a tiny local window created near a corner point to analyze features in the vicinity and filter out corner points that meet the characteristics of a picking point. In this algorithm, a square of 60 pixels is selected as the side length of the Mini-Window. After in-depth research on the corner points at the picking point, it was found that corner points meeting the requirements of a picking point have two common characteristics: First, if the Mini-Window is separated from the vertical axis of symmetry, the white area is mostly concentrated in the upper half, while the lower half is a black area; second, the upper half of the Mini-Window exhibits a "U" shape, which has three possible orientations: tilted to the left, tilted to the right, and centered. Based on the above neighborhood feature constraints, the coordinates of the picking point of the daylily plant are finally determined by filtering the local windows of candidate corner points based on geometric and grayscale distribution features.

9. The adaptive daylily segmentation and four-dimensional picking point localization method based on heterogeneous dual-domain collaboration according to claim 1, characterized in that, The algorithm for determining the daylily cutting angle based on Mini-Window, as described in step three, is as follows: First, using the Mini-Window at the picking point, extract the two intersection points where the white area intersects with the Mini-Window boundary. , ,remember for and The midpoint. Then connect with a straight line. and This line segment represents the posture of the daylily plant corresponding to this picking point, and is denoted as... Finally, draw a line perpendicular to... The line segment is denoted as Through calculation The angle between the line segment perpendicular to the right and the line segment perpendicular to the right determines the final cutting angle at the picking point. The mathematical expression for the specific calculation is as follows: ; ； ； ； ； ； in, and The function is used to obtain the left and right intersection points of the white area and the boundary of the Mini_Window. Functions used for drawing connections and A straight line, The function is used to draw the perpendicular line segment of a given line segment. The function is used to calculate the angle between a line segment and a horizontal line segment to the right.

10. The adaptive daylily segmentation and four-dimensional picking point localization method based on heterogeneous dual-domain collaboration according to claim 1, characterized in that, The analytical algorithm for the robotic arm's kinematics described in step four is as follows: Based on hand-eye calibration, the coordinates of the 3D picking point in the depth camera coordinate system are converted into the coordinates of the robot arm's world coordinate system. The inverse kinematics solution of the six-degree-of-freedom robot arm is simplified by geometric analysis, omitting redundant calculations, and the rotation parameters of the five joints other than the opening and closing of the robotic claw are obtained analytically. The cutting angle is converted into the rotation pulse value of the robotic claw through linear mapping, and the complete control command is generated by combining the joint rotation pulse values ​​to achieve precise positioning and attitude control of the robot arm.

Images