Field intensive daylily pixel classification and picking information acquisition method

Abstract

In order to solve the problems that day lily plants are densely shielded and categories are difficult to distinguish, the invention provides a field dense day lily pixel classification and picking information acquisition method. The method comprises the steps of designing a double-branch decoder fusing local and global information, constructing a progressive stepped feature mining module (PLFM) and a multi-scale hierarchical category enhancement module (PCEN), optimizing hyper-parameters through a PALA algorithm, and constructing an adaptive double-branch loss enhancement and category enhancement network (DLCE-Net) to improve semantic segmentation performance. And designing a picking point positioning algorithm (MW-GCA) of Mini-Window guided angular point analysis based on a segmentation result, screening angular point coordinates through a U-shaped window, obtaining an attitude line segment, and realizing picking point pixel coordinate and angle estimation. The method can reduce the labor loss and accelerate the engineering landing of the automatic crop picking technology.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning and intelligent crop picking, and specifically provides a method for classifying pixels of dense daylilies in a field and acquiring picking information. Background Art

[0002] Daylily, a perennial herb belonging to the Narcissus family, is widely cultivated in East Asia. Rich in nutrients, including protein, vitamins, and flavonoids, it boasts pharmacological activities encompassing multiple targets, including antiemetic, anti-inflammatory, diuretic, antidepressant, and central nervous system sedative effects. Recent studies have further revealed its efficacy in improving sleep quality. China, the world's largest producer of daylily, produces an annual output of 580,000 tons, with approximately 304,000 hectares cultivated, accounting for 97% of the world's total. However, significant technical bottlenecks exist in the harvesting process within the industry chain. Daylily harvesting is highly seasonal, typically occurring between June and August, a period of high temperatures and abundant rainfall, creating dual constraints in both time and space. Industry surveys indicate that delayed harvesting results in losses of up to 80% due to quality degradation. Globally, this loss of potential sales due to delayed harvesting amounts to $30 billion annually, severely limiting the economic benefits of the industry. Therefore, this technical dilemma highlights the urgency of research on harvesting machinery automation; Accurate segmentation of daylilies and precise location of picking points are crucial prerequisites for implementing automated harvesting systems. In recent years, with the iterative development of computer vision technology, deep learning-based techniques have provided a new path for agricultural automation. Traditional image processing methods primarily include three types of feature extraction: statistical model-based, artificial neural network-based, and knowledge-based. While statistical model-based and knowledge-based methods demonstrate robustness in specific applications, they exhibit significant limitations in daylily picking scenarios characterized by dense occlusion, high neck-leaf similarity, and difficulty distinguishing between different categories, making them difficult to achieve the required accuracy for practical applications. In contrast, neural network-based methods, leveraging the multi-level feature abstraction mechanism of convolutional neural networks, achieve collaborative parsing from pixel-level local representation to scene-level global semantics, demonstrating robustness even in complex scenarios. Li et al. proposed an improved YOLO11-GS model, integrating dynamic convolution, Ghost, BiFPN, and SimAM modules into the YOLO11n-OBB architecture, replacing Bottleneck in C3k2 to improve feature extraction performance. At the same time, they introduced oriented bounding boxes and geometric methods to achieve accurate positioning of grape stem picking points and postures, and used ByteTrack and BoT-SORT to complete dynamic tracking, providing technical support for automated picking. Rong et al. used an improved semantic segmentation model and integrated the SeMask module into the Swin Transformer. V2 encoder to realize the integration of image semantic information, and use UPerNet decoder to fuse feature selection and alignment modules, segment tomato fruit, calyx and stem areas through connected component labeling and morphological processing, and further locate the side stem picking point; Li et al. adopted an improved UNet architecture, introduced depthwise separable convolution in the encoding stage, and integrated spatial attention mechanism in the decoding stage to achieve accurate segmentation of kiwifruit and its stem. At the same time, by establishing the position constraints of fruit and stem, the target stem was locked, and the minimum bounding rectangle centroid was extracted to determine the coordinates of the picking point; Zhang et al. adopted an improved SBP-YOLOv8s-seg network, integrated self-calibration The accurate module achieves accurate segmentation of safflower filaments and fruit balls, and obtains the main feature vector of filaments through PCA, ROI extraction and contour fitting. Further combining the positional relationship between filaments and fruit balls, a staged picking point positioning algorithm is developed to effectively solve the problem of invisible picking points due to neck closure; Wang et al. used the improved Litchi-YOSO model, integrated Fasternet, C2F, Biformer and RepGFPN modules to build an efficient feature extraction framework, constructed a "point-line" model through the branch morphology reconstruction algorithm, achieved accurate segmentation of litchi, and used Prim, longest root path and Harris tests to determine the MFBB picking point;Zhu et al. used an improved YOLOv8-DWF network, integrated DCN and WIou v3 modules to enhance the detection of fresh grapes and stems, and added a small target detection layer. Based on the detection results, the geometric positioning algorithm was used to classify vertical and curved stems to achieve fast and accurate positioning of the picking point; Li et al. proposed a method combining the improved YOLOv5s and DeepLabv3+ models to achieve three-dimensional positioning of longan fruit picking points in complex natural environments by fusing depth image information. The method first used the improved YOLOv5s model to quickly detect fruit bunches and main fruit branches, and then used the improved DeepLabv3+ model to perform semantic segmentation on the main fruit branches to obtain two-dimensional coordinate information. Finally, combined with the growth characteristics of longan fruit bunches, the RGB-D information was integrated to determine the central axis and posture information of the main fruit branches, and the three-dimensional coordinates of the picking point were calculated to provide target position information for longan picking drones. Information; Zhang et al. proposed the YOLOMS multi-task learning method, which improved the detection accuracy of mangoes and fruit stems in complex environments by optimizing the backbone network of YOLOv5s and introducing the Focal-EIoU loss function. The mango stem segmentation subtask was added to the improved YOLOv5s to obtain the position and semantic information of the fruit stem. Finally, the main fruit stem recognition and picking point positioning strategy were developed to achieve the picking point positioning of clustered mangoes; Chen et al. proposed a tea bud picking point positioning method based on YOLO-v3, semantic segmentation, skeleton extraction and minimum bounding rectangle, and designed an intelligent tea end effector controlled by a personal computer and a microcontroller to deal with problems such as complex shadows and easy damage during the picking process, thereby improving the picking efficiency of the system; Yan et al. proposed an improved Mask The R-CNN model (MR3P-TS) enhances the segmentation of tea buds by extending the mask branch and introducing a multi-scale feature fusion module. The model first determines the main area of ​​the picking point through multi-connected domain analysis, then uses a minimum bounding rectangle fitting algorithm to locate the target's main axis, ultimately achieving accurate calculation of the tea bud picking point coordinates. In summary, in the context of computer technology deeply empowering agricultural automation, there are still significant technical bottlenecks in the research on picking point positioning for complex scenes such as daylily. The core difficulty is reflected in the following: the high-density clustering characteristics of daylily in the natural environment, the morphological heterogeneity of plants at the same maturity period, and the similarity of appearance at different maturity stages are intertwined, resulting in traditional visual models facing the problem of feature aliasing and cross-category semantic discrimination ambiguity in dense scenes when representing features. This not only increases the complexity of deep learning model architecture design, but also causes positioning deviations due to insufficient feature differentiation capabilities. At the same time, in high-density planting environments, the growth postures of plants are disordered, leaves and flowers are generally blocked, and the multi-source visual interference formed by superimposed morphological differences and color diversity not only seriously blocks or weakens the key positioning features of the target plants, but also causes overlapping and confusion of the visual features of different densely grown plants, making it difficult for traditional algorithms to capture effective features and unable to meet the requirements of actual operations for accuracy and stability. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the present invention aims to provide a method for classifying pixels of dense daylilies in the field and obtaining harvesting information, which is mainly used for automated harvesting of crops in real-world scenarios. To achieve the above object, the present invention provides the following technical solutions: A method for classifying dense daylily pixels in a field and acquiring harvesting information is proposed. The method constructs an ADLCE-Net dual-branch self-optimizing segmentation model, which includes a dual-branch decoder that fuses local and global information, a PCEN multi-scale parallel category focusing module, and a PLFM multi-resolution multi-receptive field progressive ladder-type feature mining module. The method uses an AUX Loss weighted auxiliary loss to guide the feature recovery process in the local decoding branch, PALA is used to optimize the high-dimensional hyperparameters of DLCE-Net to improve its reliability, and the MW-GCA algorithm is used to acquire daylily harvesting information. The dual-branch decoder includes a 3C local feature decoding module, an upsampling module, and an auxiliary loss; the PCEN includes a residual structure, a multi-scale pooling category information extraction module, and an upsampling module; the PLFM includes strip pooling, dilated convolutions with different dilation rates, and a CBR module; the MW-GCA includes a Mini-Window local window mechanism, a VASU-based "U"-shaped feature window screening algorithm, and a cropping angle calculation module; Further preferably, an ADLCE-Net multi-task self-optimizing segmentation model is constructed, and the specific steps are as follows: First, in the daylily picking scene, the spatial distribution of plants is staggered, resulting in frequent overlap and occlusion, which makes it easy to cause confusion and semantic misjudgment during feature extraction, and key features are often obscured by irrelevant information such as leaves. The present invention constructs a five-stage feature extraction backbone network Backbone to accept daylily images; the first four stages are composed of a downsampling module and a repeatedly stacked Efficient ViM module: after the image size is scaled by the convolution and regularization operations of the downsampling module, the Efficient ViM module's adaptive feature extraction capability is used to analyze the deep semantic information of the daylily plant at different scales layer by layer, providing a multi-dimensional image semantic representation for the subsequent dual-branch decoding module; and in the fifth stage, by combining the upsampling module (bilinear interpolation technology) with the repeatedly stacked Efficient ViM module, the feature map size is restored and its high-level semantic information is further deeply mined. This process provides the PCEN category information extraction module with global semantic information of the image, thereby completing the feature extraction process of the backbone network; Then, a dual-branch decoding structure with complementary local and global information is constructed, which includes two decoding branches: the global context branch and the local detail branch. In the local detail decoding branch, the 3C module is designed to achieve effective fusion of multi-scale features of daylily, and then the reconstruction of occluded edge details is completed to ensure the complete recovery of fine structures within the class. The global context branch cleverly integrates the EfficientViM Block module carried by the encoder. First, it restores the spatial size of the feature map through upsampling operations, and then uses 1×1 convolution to complete cross-channel information interaction. Finally, it is fused with the feature map output by the local detail branch to construct a global context representation that includes the overall structural relationship and relative position information of the plant, thereby forming a hierarchical and rich multi-dimensional feature representation; Next, a weighted auxiliary loss (AUX Loss) was designed to provide semantic supervision for the feature recovery process of the local detail decoding branch, promoting the refined recovery of local detail features. Furthermore, a PALA optimization strategy was designed to adaptively search for hyperparameters of the DLCE-Net model. By dynamically adjusting key parameters during network training, a self-optimizing ADLCE-Net segmentation model was formed. This strategy effectively enhanced the model's adaptability to the complex semantic structure of daylily images, further improving the model's reliability and robustness in segmentation tasks. Finally, based on the segmentation results, the MW-GCA daylily picking point intelligent positioning algorithm was further designed. By creating a "U"-shaped local window to filter corner point information, a posture line positioning method was further proposed based on the window, ultimately achieving the estimation of the pixel coordinates and angles of the picking point. The technical solution provided by the present invention has the following advantages compared with the prior art:

[0004] First, to address the accuracy issues caused by dense occlusion, large size differences within classes, and high similarity of features between classes in the segmentation of daylily images in dense scenes, the present invention proposes a DLCE-Net model. By constructing a detail-global complementary dual-branch decoder architecture and innovatively designing PLFM and PCEN feature modules, the model's perception of dense and complex features and its ability to discriminate categories are enhanced. The dual-branch architecture combines feature modules to achieve complementary fusion of multi-scale features, providing highly discriminative feature expressions with both fine-grained details and high-level semantic consistency for pixel-level positioning of picking points, effectively breaking through the bottleneck of segmentation and positioning accuracy in complex field environments; Second, to address the issues of information loss and single feature scale in target feature extraction in dense scenes, making it difficult to obtain deep and refined features, this paper proposes a dense-scale progressive ladder feature mining module (PLFM). This module uses a deep coupling architecture of strip pooling, dilated convolution with multiple dilation rates, and a CBR module to construct a multi-resolution feature pyramid with hierarchical progressive characteristics. It achieves refined analysis of dense-scale feature space through a gradient receptive field regulation mechanism, significantly enhancing the model's spatial perception of daylily plants in complex planting environments. Third, to address the problem of feature discreteness and category confusion in daylily maturity segmentation, which leads to a decrease in model segmentation accuracy, this paper proposes a multi-scale hierarchical category enhancement module (PCEN). This module integrates residual, multi-scale pooling, and upsampling structures. Based on a multi-scale pooling strategy, it hierarchically samples high-level semantic features, constructs a hierarchical feature aggregation structure, extracts key discriminative features, suppresses redundant information, and effectively improves plant segmentation performance in dense daylily planting scenarios. Fourth, to address the problems of inefficient feature utilization and severe semantic loss in densely occluded daylily picking scenes using a single-branch decoder, this paper constructs a dual-branch decoder that integrates local and global information. By leveraging the complementary advantages of local details and global context branches, it solves the problems of detail loss and feature confusion in densely occluded scenes. Fourth, to address the problems of inefficient feature utilization and severe semantic loss in densely occluded daylily picking scenes using a single-branch decoder, this paper constructs a dual-branch decoder that integrates local and global information. By leveraging the complementary advantages of local details and global context branches, it solves the problems of detail loss and feature confusion in densely occluded scenes. Fifth, to address the difficulty of implementing differentiated supervision on the daylily plant feature processing output by the multi-scale feature extractor when using a single loss function in a dual-branch decoder architecture, this paper proposes a primary-auxiliary collaborative loss function (MACLF). The primary loss function strengthens global semantic constraints to address large-scale contour segmentation, while the auxiliary loss function generates refined supervision signals to optimize densely occluded areas and small-scale details, achieving differentiated feature constraints in the dual-branch decoder and improving segmentation accuracy and picking point positioning robustness. Sixth, in response to the bottleneck problems of the Artificial Lemming Algorithm (ALA) in terms of convergence efficiency, global optimization ability and adaptability to complex constraints, the present invention proposes a parallelized Artificial Lemming Optimization Algorithm (PALA) and constructs a hyperparameter self-optimization network model ADLCE-Net. PALA uses the synergistic effect of a multi-subpopulation parallel search strategy and a diversity preservation mechanism, takes the model mean intersection over union (mIoU) as the optimization target, sets 14 key hyperparameters as decision variables, and realizes efficient hyperparameter adaptive optimization in complex high-dimensional spaces. This method improves global exploration efficiency while ensuring population diversity by decoupling the limitations of traditional single-population search, ultimately enhancing the robustness and reliability of the DLCE-Net model under complex constraints, thereby achieving accurate pixel-level semantic segmentation performance; Seventh, to address the problem that multi-source visual interference in actual planting environments causes key positioning features of plants to be obscured and confused, and traditional algorithms are unable to meet the accuracy requirements of picking operations, the present invention proposes the MW-GCA daylily picking point positioning algorithm. This algorithm relies on the segmentation results of the ADLCE-Net model to screen out mature and over-mature daylily mask images, constructs a corner point extraction strategy to obtain potential picking point locations, and designs a Mini-Window local window mechanism to screen corner point coordinates that meet the characteristics, thereby achieving pixel-level coordinate acquisition of daylily picking points in complex environments; Eighth, in response to the technical need to reduce plant damage rate and maintain plant appearance integrity to improve economic benefits due to the impact of cutting angle on picking quality during automated daylily harvesting, the present invention proposes a method for determining the cutting angle of daylily plants. This method creates a Mini-Window local window at the picking point, constructs a plant end posture line segment, and then solves the perpendicular line of this line segment to determine the optimal cutting angle. This method aims to provide precise angle parameters for the robotic claw picking operation, reduce the risk of plant damage caused by angle deviation, and ensure the integrity of the harvest. The technical architecture of this invention is rationally designed and is suitable for tasks such as splitting daylilies and locating picking points. By replacing manual labor with automation, manpower input can be significantly reduced and operational efficiency improved. This not only aligns with the national strategy for independent innovation in intelligent manufacturing, but also accelerates the intelligent transformation of the industry and promotes the establishment of a sustainable development model for the daylily industry. Given its technical practicality and industrial enabling value, this invention has broad market application space and commercial potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0005] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present invention and, together with the description, serve to explain the principles of the present invention; To clearly present the technical solutions of the embodiments of the present invention or the prior art, the following briefly describes the drawings involved in the description of the embodiments or the analysis of the prior art. A person skilled in the art can reasonably deduce other relevant drawings based on the contents of the above drawings without inventive effort. Figure 1 It is the overall flow chart of the present invention; Figure 2 This is the DLCE-Net model architecture diagram; Figure 3 This is the 3C module structure diagram; Figure 4 This is the PLFM module structure diagram; Figure 5 This is the PCEN module structure diagram; Figure 6 This is the PALA algorithm flow chart; Figure 7 This is the flow chart of the MW-GCA picking point positioning algorithm; Figure 8 This is the flow chart of the algorithm for determining the cropping angle of daylily based on Mini-Window; Figure 9 This is a visualization diagram of the comparative experimental results; Figure 10 Schematic diagram of segmentation effects of different models. DETAILED DESCRIPTION

[0006] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the scheme of the present invention will be further described below; The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings; The present invention proposes an integrated solution AD-MDNet for intelligent daylily picking with the dual-branch decoding semantic segmentation model ADLCE-Net as the core. The ADLCE-Net model constructs a dual-branch decoder with complementary local and global information, a PLFM progressive ladder feature mining module, and a PCEN multi-scale hierarchical category enhancement module, aiming to effectively solve the key problems affecting segmentation accuracy in dense daylily image segmentation, such as dense occlusion, significant size differences within classes, and high similarity of features between classes. By designing the Main-Auxiliary Collaborative Loss Function (MACLF), the weight proportion of the auxiliary loss (AUX Loss) in the total loss is dynamically adjusted, highlighting the dominant role of the main loss function in the training process; at the same time, the PALA hyperparameter self-optimization strategy is designed to adaptively search and optimize the model hyperparameters, significantly improving the model's self-learning ability in dense daylily picking scenarios. Based on the precise segmentation results of ADLCE-Net, the MW-GCA daylily picking point positioning and cropping angle calculation method is further constructed. This method innovatively combines the corner point extraction strategy with the Mini-Window local window corner point screening mechanism to achieve efficient screening of corner point coordinates that meet the picking point characteristics. It further obtains the posture line segment of the plant end based on local window analysis, and then accurately calculates the cutting angle, completing the intelligent acquisition of key information for daylily picking. A method for classifying pixels of dense daylilies in the field and obtaining harvesting information, such as Figure 1 、 2 As shown in the figure, the ADLCE-Net dual-branch self-optimizing segmentation model is constructed. The ADLCE-Net dual-branch self-optimizing segmentation model includes a dual-branch decoding architecture, a progressive step-wise feature mining module (PLFM), and a multi-scale hierarchical category enhancement module (PCEN). The main-auxiliary collaborative loss strategy (MACLF) is used to ensure the model's feature recovery capability, the PALA high-dimensional hyperparameter optimization algorithm is used to improve the model's reliability, and the MW-GCA is used to determine the pixel coordinates of the daylily picking point and the cropping angle information. Specifically, the dual-branch decoder includes a 3C local feature decoding module, an upsampling module, and an auxiliary loss; the PCEN includes a residual structure, a multi-scale pooling category information extraction module, and an upsampling module; the PLFM includes strip pooling, dilated convolutions with different dilation rates, and a CBR module; and the MW-GCA includes a Mini-Window local window mechanism, a "U"-shaped feature window screening algorithm based on VASU, and a rotation angle calculation module.

[0007] First, we build the ADLCE-Net dual-branch self-optimizing segmentation model; First, in the daylily picking scene, the spatial distribution of plants is staggered, resulting in frequent overlap and occlusion, which makes it easy to cause confusion and semantic misjudgment during feature extraction, and key features are often obscured by irrelevant information such as leaves. The present invention constructs a five-stage feature extraction backbone network Backbone, which accepts 512×512 daylily images; the first four stages are composed of a downsampling module and a repeatedly stacked Efficient ViM module: after the image size is scaled by the convolution and regularization operations of the downsampling module, the Efficient ViM module's adaptive feature extraction capability is used to analyze the deep semantic information of the daylily plant at different scales layer by layer, providing a multi-dimensional image semantic representation for the subsequent dual-branch decoding module; in the fifth stage, by combining the upsampling module (bilinear interpolation technology) with the repeatedly stacked Efficient ViM module, the feature map size is restored and its high-level semantic information is further mined. This process provides the PCEN category information extraction module with global semantic information of the image, thus completing the feature extraction process of the backbone network; Then, a dual-branch decoding structure with complementary local and global information is constructed, which includes two decoding branches: the global context branch and the local detail branch. In the local detail decoding branch, the 3C module is designed to achieve effective fusion of multi-scale features of daylily, and then the reconstruction of occluded edge details is completed to ensure the complete recovery of fine structures within the class. The global context branch cleverly integrates the EfficientViM Block module carried by the encoder. First, it restores the spatial size of the feature map through upsampling operations, and then uses 1×1 convolution to complete cross-channel information interaction. Finally, it is fused with the feature map output by the local detail branch to construct a global context representation that includes the overall structural relationship and relative position information of the plant, thereby forming a hierarchical and rich multi-dimensional feature representation; Finally, a weighted auxiliary loss (AUX Loss) was designed to provide semantic supervision for the feature recovery process in the local detail decoding branch, promoting the refined recovery of local detail features. Furthermore, a PALA optimization strategy was designed to adaptively search for hyperparameters of the DLCE-Net model. By dynamically adjusting key parameters during network training, a self-optimizing ADLCE-Net segmentation model was developed. This strategy effectively enhanced the model's adaptability to the complex semantic structure of daylily images, further improving its reliability and robustness in segmentation tasks.

[0008] Second, construct a dual-branch decoder; In order to solve the problem of low feature utilization and semantic information loss of the single-branch decoder in the densely occluded daylily picking scene, the present invention proposes a dual-branch decoder structure: the local detail branch reconstructs the edge details of the densely occluded area through upsampling and feature splicing, and restores the detailed information of the complex morphology of similar plants; the global context branch captures long-distance dependencies through EfficientViM Block, constructs a global representation including the overall structure and relative position of the plant, and integrates the edge, texture and other information of the local detail branch to solve the problem of feature confusion and high similarity between classes, and form a contextual representation with both details and global semantics. In terms of the local detail branch, the encoding result of the feature map will first enter the 3C (Cat+Channel Attention+CBR) module (the 3C module structure is as follows Figure 3 This module achieves pixel-level detail recovery by fusing feature information from daylily images at different scales. CBR stands for Conv+BatchNorm+ReLU. The 3C module fuses feature information from daylilies at different scales in the following manner: first, it performs feature alignment and concatenation; second, it applies Channel Attention filtering to the fused feature map to remove feature information from useless channels; finally, the CBR module adjusts the number of channels in the feature map to restore it to the original input channel count. The specific mathematical expression for this calculation is as follows: ; ; ; ; ; ; in, and Respectively represent the output results of the encoder of the current stage and the stage after the current stage; Indicates that the feature map and Splice by channel; and represents a fully connected layer with shared parameters; Represents the Sigmoid activation function; express Convolution kernel; In terms of global context branches, the present invention constructs a global decoding structure based on EfficientViM Block, which is combined with the local detail branch through upsampling operation and 1x1 convolution. By fusion, we can achieve global distribution perception of different feature images of daylily, and thus realize global feature recovery. The specific mathematical expression is as follows: ; ; Among them, when hour, Represents the output result of the previous global decoding stage, which ; Especially, when hour, After decoding in the EfficientViM Block module, the global context decoding branch The output of the decoding stage The specific mathematical expression of the calculation is shown above. express Convolution kernel. Finally, the final output of the global context decoder is The daylily plant category information extracted by the PCEN module Perform adaptive fusion to generate the final semantic segmentation result. The specific mathematical expression is as follows, where: represents the learnable adaptive fusion parameters; ; .

[0009] Third, construct PLFM; In semantic segmentation tasks, the edge information of the target is crucial for defining boundaries and enhancing differentiation. However, in dense daylily planting scenes, the overlap and occlusion of plants are significant, making it difficult to define boundaries, which in turn causes segmentation errors and reduces segmentation accuracy. To solve this problem, the present invention proposes a progressive ladder feature mining module (PLFM), such as Figure 4 This module aims to mine key feature information that is often overlooked due to dense occlusion of plants in dense-scale scenes. Through a multi-receptive field and multi-resolution progressive ladder-type feature mining method, it achieves refined extraction of deep feature information in dense-scale scenes, thereby enhancing the model's perception of dense-scale scenes. Specifically, the PLFM module consists of three similar feature extraction submodules. Each submodule first processes the input features through dilated convolutions with different dilation rates. Since convolutions with different dilation rates have different receptive fields, they can extract feature information of different resolutions under dense scales. Then, the processed features are sampled through the pooling layer to further extract important information and reduce the interference of redundant information on model training. During the pooling process, the present invention constructs a pooling system including three types: maximum pooling (MaxPool), average pooling (AvgPool) and strip pooling (StripPool). Among them, strip pooling can efficiently capture long-range dependencies while also processing strip-shaped targets well, which is very suitable for daylily segmentation tasks. By stacking these three submodules and constructing a residual connection method, the PLFM module can effectively enhance the model's perception of the characteristics of daylily plants in densely occluded scenes when the daylily plants are densely intertwined and mutually occluded, thereby improving the segmentation accuracy. For the input of a certain stage of the encoder , the output after processing by the PLFM module The specific mathematical expression is as follows: ; ; ; ; ; ; ; in, represents the dilated convolution with a dilation rate of r, refers to the ReLU activation function, and Represents maximum pooling and average pooling respectively, Stands for strip pooling.

[0010] Fourth, build PCEN; In the task of distinguishing the maturity of daylily, the significant morphological differences between individual plants of the same maturity lead to discrete feature distribution within the category, while the highly similar visual features between different maturity categories cause feature confusion between categories. These two core problems make it easy for the segmentation model to have blurred regional boundary divisions and misjudgment of category labels when dealing with mutually occluded plants, which significantly affects the segmentation accuracy. To this end, the present invention proposes a multi-scale hierarchical category enhancement module (PCEN). The design of this module is based on the advantages of the multi-scale pooling strategy in category differentiation tasks and its adaptability to the task of extracting category information of daylily. By configuring multi-scale pooling kernels to perform hierarchical sampling of high-level semantic features, PCEN constructs a hierarchical feature aggregation structure with multi-scale context-awareness to achieve hierarchical extraction of key category discriminant features. This multi-dimensional feature extraction mechanism not only effectively addresses the problem of discrete feature distribution caused by morphological differences among similar plants, but also enhances the model's ability to distinguish subtle differences between different maturity categories by adaptively suppressing irrelevant redundant information, thereby significantly improving the segmentation performance of plants in densely occluded scenes. Its specific structure is as follows. Figure 5 As shown; Specifically, for the high-level features of the input First, it is input into the multi-scale pooling module, and three pooling kernels with kernel sizes of 3×3, 7×7 and 11×11 are set for feature extraction. For each pooling kernel, the maximum pooling (MaxPool) and average pooling (AvgPool) operations are performed respectively, and the advantages of the two pooling methods are combined to obtain a richer multi-scale feature representation. Subsequently, the extracted multi-scale features are subjected to residual fusion to retain the image detail information and contextual semantic association. Finally, after a 2x upsampling operation and 1×1 convolution processing, the spatial alignment of high-level semantic information and the output features of the dual-branch decoder is achieved, providing standardized input for the subsequent feature fusion with the segmentation head. The output of the module PCEN The mathematical expression for the calculation is defined as follows, where Indicates a 2x upsampling operation, and The corresponding kernel sizes are Maximum pooling and average pooling; ; ; .

[0011] Fifth, construct the main and auxiliary collaborative loss function; In the semantic segmentation task of daylily images, when adopting a dual-branch decoder architecture and using a single loss function, the model struggles to implement effective differentiated supervision of the processing of daylily plant feature information extracted by the multi-scale feature extractor. This lack of supervision is primarily manifested in the following: the decoding branch responsible for local detail features lacks precise semantic guidance and is unable to effectively distinguish and utilize feature information at different scales when fusing multi-scale features, resulting in inaccurate recovery of daylily plant details and, in turn, affecting overall segmentation accuracy. To address this issue, the present invention constructs a primary-auxiliary collaborative loss function (MACLF), which is specifically expressed mathematically as follows: ; Specifically, the present invention constructs an auxiliary loss function (AUX Loss) in the local detail decoding branch (such as Figure 2 As shown in Figure 2), by applying refined semantic supervision signals to guide the feature refinement recovery process of this branch, the semantic consistency between multi-scale features is effectively enhanced, and the overall segmentation accuracy of the complex structure of the daylily plant is finally improved. In addition, in order to clarify the leading role of the main loss function, the present invention designs a weight parameter It is used to adjust the weight of the auxiliary loss term in the total loss function. This parameter will be determined in the subsequent DLCE-Net hyperparameter self-optimization process. and The specific mathematical calculation expression is as follows: ; ; The auxiliary loss function is the same as the main loss function of DLCE-Net, which is composed of the cross entropy loss function and Loss Function composition. and The specific expression of is as follows. Where K represents the total number of categories predicted by the segmentation model, represents the number of samples, Indicates in Among the samples, the The true label of the class, Indicates in Among the samples, the model predicts that The probability value of the class, ; ; ; The input of the auxiliary loss function is the decoding result of the last stage in the local detail decoding structure ,exist Before inputting into the auxiliary loss function, it needs to be processed. First, a 1x1 convolution layer is used to adjust the number of channels of the image to the total number of categories predicted by the segmentation model. , then, the image is restored to the input resolution through upsampling. The specific mathematical expression of the processing process is as follows, where, Represents the input value of the auxiliary loss function; .

[0012] Sixth, we designed the PALA strategy to optimize the hyperparameters of the DLCE-Net model; Aiming at the shortcomings of the ALA algorithm in convergence efficiency, global optimization and adaptability to complex constrained scenarios, especially the search imbalance problem exposed in the high-dimensional hyperparameter optimization of DLCE-Net, this paper proposes the PALA strategy suitable for self-optimization of semantic segmentation model hyperparameters. This strategy constructs a systematic optimization framework through three aspects: in the population initialization stage, a population initialization mechanism based on the good point set is designed, and its low-dimensional projection uniformity is used to improve the balanced coverage of the initial population in the search space, thereby consolidating the foundation for global exploration; in the search mechanism optimization, a multi-subpopulation collaborative system including exploration, development, and balance types is designed, and through differentiated parameter configuration (the exploration type adopts large perturbations and low convergence coefficients, the development type adopts small perturbations and high convergence coefficients, and the balance type adopts medium parameters), search space decoupling and complementary exploration are achieved, and a ring neighborhood optimal solution sharing mechanism is provided. A cross-population individual replacement strategy with a ratio of 10% (rounded down and ≥1) is used to promote the diffusion of high-quality solutions and maintain population diversity; when dealing with the problem of premature convergence, an elite reverse learning strategy is integrated to generate reverse solutions by screening the top 10% fitness individuals to construct a reverse population, and generate offspring individuals through a preferential selection mechanism, deeply mining elite information to enhance population diversity. The above improvements, through the synergy of multi-subpopulation parallel search and diversity preservation mechanism, effectively overcome the limitation of the original algorithm that is prone to falling into local optimality, and significantly enhance the global exploration capability in complex hyperparameter spaces; In the process of constructing the semantic segmentation model DLCE-Net, the present invention establishes DLCE-Net as the core architecture and designs the PALA algorithm as a hyperparameter optimization strategy, through which the model hyperparameters are systematically optimized. During the optimization process, 14 key hyperparameters, including the optimizer type (sgd, adam), the initial learning rate size (lr), the momentum coefficient (momentum), the learning rate reduction method (step, cos), the auxiliary loss weight coefficient (λ), and the 9 expansion rates (Expand_ratio) of the HSM-SSD module in the encoder and decoder, are used as decision variables of the PALA algorithm. In the optimization process of the PALA algorithm, each hyperparameter combination of DLCE-Net is regarded as an individual, and different hyperparameter combinations constitute the initial population of the PALA algorithm. Through the adaptive adjustment of these hyperparameters by the PALA algorithm, the present invention constructs the final ADLCE-Net model. In the PALA algorithm, the fitness function is used to evaluate the quality of each individual, thereby guiding the algorithm's search process. The mathematical expression of the specific calculation of the fitness function defined for the DLCE-Net hyperparameter search in the present invention is as follows: ; ; in, The function indicates that the model training is obtained under the specified 14-dimensional hyperparameter setting. value, Represents the fitness function. After searching for the optimal hyperparameters using the PALA optimization strategy, the present invention successfully obtained the optimal hyperparameter combination of the model and finally used it as the hyperparameter setting of ADLCE-Net, as shown in Table 1; Table 1 PALA adaptive optimization results of DLCE-Net hyperparameters Hyperparameter name Search Scope Search for optimal content Search results Optimal results optimizer (0,2) {adam, sgd} 0.65143 adam lr (1e-5, 1e-2) (1e-5, 1e-2) 0.00051 0.00051 momentum (0.6, 0.95) (0.6, 0.95) 0.88322 0.88322 lr_way (0,2) {step, cos} 1.15946 cos λ (0,5) {0.2, 0.4,0.6,0.8,1.0} 3.68312 0.8 Expand_ratio[0] (0,3) {16,32,64} 2.34783 64 Expand_ratio[1] (0,3) {16,32,64} 0.86806 16 Expand_ratio[2] (0,3) {16,32,64} 2.48956 64 Expand_ratio[3] (0,3) {16,32,64} 2.65453 64 Expand_ratio[4] (0,3) {16,32,64} 1.89349 32 Expand_ratio[5] (0,3) {16,32,64} 0.24635 16 Expand_ratio[6] (0,3) {16,32,64} 0.91811 16 Expand_ratio[7] (0,3) {16,32,64} 2.46358 64 Expand_ratio[8] (0,3) {16,32,64} 2.32776 64 Among them, Expand_ratio[0]-Expand_ratio[4] represent the five expansion ratios of HSM-SSD in the five stages of the encoder, and Expand_ratio[5]-Expand_ratio[8] represent the four expansion ratios of HSM-SSD in the four stages of the decoder. When processing the discretized optimization range, the present invention determines the final search value by the index number of the array, and rounds down the values ​​of the corresponding dimensions of the PALA search solution vector to obtain the final optimization result. From the data in Table 1, it can be seen that the PALA algorithm successfully searches for the optimal combination of 14-dimensional hyperparameters. The present invention applies this hyperparameter combination to DLCE-Net to form the final model ADLCE-Net, further improving its performance in the segmentation task.

[0013] Seventh, design the MW-GCA daylily picking point positioning algorithm; In high-density planting of daylilies, the plant posture and terminal morphology are complex, with superimposed leaf occlusion, background interference, stem bending, and variable flower morphology and color, making it difficult for traditional picking point positioning algorithms to ensure robustness. To address the above problems, this algorithm combines corner point extraction strategies with an innovatively designed Mini-Window mechanism to achieve the acquisition of picking information in complex environments. Specifically, the mask merged image of the mature and immature areas is extracted. , and then obtain The geometric coordinates of the corner points of the image, the specific mathematical expression is as follows. In order to solve the problem of reduced positioning accuracy due to insufficient integrity of the edge area mask, the boundary corner points are optimized. Then, the potential picking points are located through the corner detection strategy, and the corner filtering strategy is implemented for each connected area, focusing on eliminating invalid corner points at the top of the plant, and focusing the analysis on the characteristic significant area at the bottom of the plant. The local feature analysis window is constructed through the innovative Mini-Window mechanism, which can effectively block irrelevant background interference outside the window, so that the algorithm can focus on the key features of the target area (the algorithm process is as follows Figure 7 shown); ; Specifically, the Mini-Window refers to a tiny local window created near a corner point, used to analyze features near the corner point and select corner points that meet the requirements for picking points. In this algorithm, the present invention selected a 60-pixel square as the side length of the Mini-Window. After in-depth research on the corner points at the picking points, it was found that the corner points that meet the picking point requirements all 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 black; second, the upper half of the Mini-Window exhibits a "U"-shaped feature, and this "U"-shaped area can be tilted to the left, tilted to the right, and centered. Based on the above neighborhood feature constraints, the local window of the candidate corner points is screened for geometric and grayscale distribution features to ultimately determine the precise picking point coordinates of the daylily plant.

[0014] Eighth, design a Mini-Window-based algorithm for determining the cropping angle of daylily; During the daylily picking process, the selection of cutting angles plays a crucial role in the quality of the harvest. A reasonable cutting angle not only effectively reduces damage to the plants but also ensures that the picked daylilies are in good condition, thereby improving the overall economic benefits of automated harvesting. To address this issue, the present invention conducts an in-depth analysis of the Mini-Window at the target corner point in the Mini-Window and constructs a cutting angle calculation method based on the Mini-Window. Specifically, first, by picking the Mini-Window at the picking point, extract the two intersection points where the white block intersects with the Mini-Window boundary. 、 ,remember for and The midpoint of Figure 8 Then, connect the and , which is the posture line segment of the daylily plant corresponding to the picking point, recorded as Finally, draw a line perpendicular to The line segment is denoted as By calculation The angle between the vertical right line segment and the vertical right line segment can determine the final cutting angle of 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 block and the Mini_Window boundary. Function used to draw connections and A straight line, The function is used to draw a perpendicular line segment of a line segment. The function is used to calculate the angle between a line segment and a horizontal line segment to the right.

[0015] This study constructed a semantic segmentation dataset for daylilies grown in dense field environments called TYUST-Daylily. Using LabelMe, 700 images were annotated pixel-wise, categorizing the flowers into mature, immature, and overmature categories based on their growth stage. Data augmentation was performed using random brightness adjustment, noise injection, and horizontal / vertical flipping, increasing the data size to 1400 images. Finally, the training and validation sets were divided into a non-overlapping 8:2 ratio to provide standardized data support for the DLCE-Net model.

[0016] In order to systematically evaluate the optimization performance of the improved PALA algorithm, the present invention uses representative cutting-edge algorithms in the field of intelligent optimization in the past two years as comparison objects, including the Snow Goose Algorithm (SGA), the Puma Optimizer (PO), the Artificial Protozoa Optimizer (APO), the Fungal Growth Optimizer (FGO), the Dream Optimizer (DOA) and the original ALA algorithm. These algorithms have demonstrated excellent optimization capabilities in related studies and can provide a reliable reference for verifying the effectiveness of the PALA algorithm. At the same time, in the comparative experimental design, 13 high-dimensional test functions are used as evaluation benchmarks, including 7 high-dimensional unimodal functions and 6 high-dimensional multimodal test functions, covering typical unimodal extreme value solutions and multimodal complex optimization scenarios. The performance of the algorithm in multiple key performance dimensions such as global search capability, local fine development capability, and anti-premature convergence capability can be effectively examined. By constructing this widely representative test function set, a scientific and reasonable experimental platform is provided for the performance comparison of different algorithms, ensuring that the experimental results can objectively reflect the actual effectiveness of each algorithm in high-dimensional optimization problems. In the comparative experiments of the optimization algorithms, this paper uniformly set the dimensionality to 30, the maximum number of iterations to 500, and the population size to 100. For the PALA algorithm, this paper set the number of subpopulations to 5, the number of individuals per subpopulation to 20, and the communication gap between subpopulations to 10. The other configurations of the optimization algorithms remained consistent with those in the original paper. To further eliminate the influence of randomness, all functions in this paper were run 30 times independently. The mean (mean) and standard deviation (std) of the 30 experimental results were used as evaluation metrics. The optimal values ​​of the current test function are indicated in bold. The experimental results are shown in Tables 2 and 3. Table 2 High-dimensional single-peak simulation comparison experiment

[0017] Table 3 High-dimensional multi-peak simulation comparison experiment

[0018] The PALA algorithm demonstrated excellent performance in high-dimensional unimodal and multimodal simulation comparison experiments. In the high-dimensional unimodal experiments, PALA's mean and standard deviation values ​​for various test functions (F1-F7) far outperformed most compared algorithms. For example, in the F1 test, PALA achieved a mean value of 2.674e-38 and an std value of 4.867e-38, significantly lower than those of other algorithms, demonstrating its good convergence and strong stability in high-dimensional unimodal environments. In the high-dimensional multimodal experiments, PALA also performed well across all test functions. For example, in the F11 test, the mean value reached 0.000e+00 and the std value reached 0.000e+00, significantly outperforming other algorithms. This demonstrates that PALA not only accurately optimizes in complex high-dimensional multimodal scenarios, but also maintains high stability. Compared to other compared algorithms such as SGA, PO, and APO, PALA has a significant advantage in high-dimensional optimization capabilities.

[0019] Given that the PLFM module aims to address high-level feature mining at dense scales, the optimal placement of the module within the encoder output scale is a key focus of this research. However, due to its continuous atrous convolution mining architecture, simultaneous application to output feature mining at multiple encoding stages can lead to vanishing gradients during backpropagation, resulting in training failure. To address this issue, this paper conducted comparative experiments, evaluating the performance of the PLFM module by deploying it at the output of the first four encoding stages. To ensure the fairness of the experiments, all experiments were conducted under the same hardware environment and configuration parameters. The experiment presets the maximum number of iterations to 200 times, the batch size (batch-size) to 4, and the remaining hyperparameters are set to the results of PALA optimization. A comprehensive comparative analysis of different schemes was conducted from four aspects: mean intersection over union (mIoU), mean average precision (mAP), parameter quantity, and pixel accuracy. The specific results are shown in Table 4. The analysis found that the deployment of the PLFM module at different output scale positions of the encoder will have a certain impact on the segmentation performance. Among them, the deployment scheme at position 1 has a slightly improved segmentation performance compared with the schemes at positions 2, 3, and 4, while the parameter quantity level is relatively low. Taking into account the performance and parameter quantity, the present invention finally selects position 1 as the deployment position of the PLFM module; Table 4 Comparison of PLFM position experiments Serial number Location Params(M) MIou(%) mAP(%) Precision(%) 1 Position 1 7.894M 70.68 82.20 96.73 2 Position 2 7.991M 70.15 81.87 96.67 3 Position 3 8.378M 70.32 82.23 96.70 4 Position 4 9.920M 70.24 81.81 96.68

[0020] To verify the effectiveness of each module of the ADLCE-Net model, this paper constructed a baseline model based on the EfficientViM Block and conducted ablation experiments. Using the control variable method, under the same device environment, the maximum number of iterations was set to 200 and the batch size was set to 4. The remaining hyperparameters were set according to the PALA optimization results. The experimental results are shown in Table 5. In terms of parameter volume, when only the base and PLFM modules are enabled, the parameter volume is 5.779M, and when all modules are enabled, the parameter volume increases to 7.894M, with an average increase of approximately 0.35M for each additional module. In terms of performance, mIoU increased from 68.80 to 70.68, an increase of 2.73%; mAP increased from 80.91 to 82.20, an increase of 1.60%; and Precision increased from 96.51 to 96.73, an increase of 0.23%. Overall, the addition of each module effectively improved the performance of the ADLCE-Net model while increasing the number of parameters. In particular, the PCEN module has a significant effect in improving model accuracy, verifying the necessity and effectiveness of each module in optimizing the model. Table 5 ADLCE-Net model ablation experiment Serial number base 3C AUXLoss PCEN PCEN(weight) PLFM Params / M mIou(%) mAP(%) Precision(%) 1 ✔ ✔ 5.779M 68.80 80.91 96.51 2 ✔ ✔ ✔ 6.377M 69.29 80.86 96.58 3 ✔ ✔ ✔ ✔ 6.377M 69.82 81.19 96.64 4 ✔ ✔ ✔ ✔ ✔ 7.894M 70.06 82.37 96.64 5 ✔ ✔ ✔ ✔ ✔ 7.862M 70.25 81.84 96.70 6 ✔ ✔ ✔ ✔ ✔ ✔ 7.894M 70.68 82.20 96.73 In order to further verify the recognition effect of the modules in ADLCE-Net on three different categories in the wild dense daylily dataset, the present invention designed a category ablation experiment. The experimental results are shown in Table 6. The results show that the accuracy of the over-mature category increased from 67.57 to 69.37, an increase of 1.8, an improvement rate of 2.66%, and an average improvement of about 0.3 each time; the accuracy of the immature category increased from 52.89 to 55.03, an increase of 2.14, an improvement rate of 4.05%, and an average improvement of about 0.36 each time; the accuracy of the mature category increased from 57.68 to 61.12, an increase of 3.44, an improvement rate of 5.96%, and an average improvement of about 0.57 each time. The activation of each module improves the accuracy of each category and the overall segmentation of the model, and the improvement of the mature daylily category is more obvious, verifying the positive role of the adaptive PCEN module and other modules in improving the accuracy of category recognition; Table 6 ADLCE-Net category ablation experiment Serial number base 3C AUXLoss PCEN PCEN(weight) PLFM background overripe unripe ripe mIoU(%) 1 ✔ ✔ 97.07 67.57 52.89 57.68 68.80 2 ✔ ✔ ✔ 97.09 68.30 52.97 58.80 69.29 3 ✔ ✔ ✔ ✔ 97.13 68.67 54.08 59.40 69.82 4 ✔ ✔ ✔ ✔ ✔ 97.15 68.77 54.51 59.82 70.06 5 ✔ ✔ ✔ ✔ ✔ 97.21 69.27 54.52 60.00 70.25 6 ✔ ✔ ✔ ✔ ✔ ✔ 97.20 69.37 55.03 61.12 70.68

[0021] In order to further verify the superiority of the segmentation model proposed in this paper, we compared the advanced segmentation models of the same level in recent years, including DeepLab v3+, DDRNet-23, SegFormer-b1, SegNext-s, PIDNet-m, FSegNet and DSNet-head256. In order to ensure the fairness of the experiments, all experiments were carried out under the same hardware environment and configuration parameters, with a maximum number of iterations of 200 and a batch size of 4. The comparison method was comprehensively compared from four dimensions: mIoU, mAP, parameter quantity, and pixel accuracy. The comparison results are shown in Table 7 and Figure 9 As shown. In the segmentation model comparison experiment, the difference in the number of parameters of each model is significant. FSegNet is as high as 54.000M, while the model of the present invention is only 7.894M, showing excellent lightness. In terms of performance indicators, the model of the present invention has outstanding advantages, with mIoU of 70.68, 2.85 higher than the second highest DSNet-head256; mAP reaches 82.20, 2.12 higher than the second highest DSNet-head256 model; Precision is 96.73, 0.37 higher than the second highest DSNet-head256. This shows that the ADLCE-Net model of the present invention surpasses other models in segmentation accuracy, target recognition and prediction accuracy with a lower number of parameters, and has high efficiency and superiority in the daylily segmentation task; Table 7 Segmentation model comparison experimental results Serial number Model years Params / M mIoU(%) mAP(%) Precision(%) 1 DeepLab v3+ 2018.10 5.814M 66.89 78.90 96.27 2 DDRNet-23 2021.1 20.147M 67.35 79.02 96.21 3 SegFormer-b1 2021.5 13.678M 62.65 75.27 95.63 4 SegNext-s 2022.9 14.479M 67.39 79.99 96.20 5 PIDNet-m 2023.6 28.537M 66.33 78.02 96.12 6 FSegNet 2024.5 54.000M 62.12 74.54 95.68 7 DSNet-head256 2024.6 9.987M 67.83 80.08 96.36 8 Ours 2025 7.894M 70.68 82.20 96.73 like Figure 10 As shown in the figure, the proposed model demonstrates excellent segmentation performance in high-density, complex scenes: it accurately outlines the plant outlines and achieves efficient segmentation, clearly distinguishing different categories through color coding. Compared with the standard label image, the segmentation results are highly close to the actual annotations, maintaining high plant segmentation integrity and category differentiation even in complex backgrounds, fully demonstrating the model's robust segmentation capabilities in complex environments.

[0022] The present invention is carried out in a 14th Gen Intel (R) Core (TM) i5 - 14600KF @3.50GHz, NVIDIA GeForce RTX 4070 Ti SUPER, 32GB of memory, Linux Ubuntu 20.04 operating system and PyTorch 3.8 environment.

[0023] The present invention has a wide range of applicability and can be effectively applied to the division and picking of daylilies. In actual application, the invention can significantly reduce the economic losses caused by manual operation limited by experience level and physical strength, and effectively promote the engineering process of automated crop harvesting technology. From a macro perspective, the present invention actively conforms to the country's strategic orientation of promoting the healthy development of the intelligent manufacturing industry and encouraging independent innovation. By empowering agricultural production links through technology, it injects momentum into the intelligent transformation and upgrading of the agricultural industry and helps build a modern agricultural production system.

[0024] The above content is only a specific implementation method of the present invention, which is used to help those skilled in the art understand and implement the present invention. The scope of protection of the present invention is not limited to the enumerated embodiments. Any reasonable improvement based on the technical ideas and principles of the present invention, any equivalent replacement or embellishment optimization of local technical features, as long as they do not deviate from the essence of the core technical solution of the present invention, should be included in the protection scope defined by the claims. Adaptive adjustments made by ordinary technicians in this field without violating the principles of the invention also fall within the scope of protection of the present invention.

Claims

1. A method for classifying pixels of densely populated daylilies in a field and obtaining harvesting information, characterized by: Construct a dual-branch decoding semantic segmentation model ADLCE-Net, which includes a dual-branch decoder, a PCEN multi-scale parallel category focusing module, and a PLFM multi-resolution multi-receptive field progressive step-by-step feature mining module; The AUX Loss weighted auxiliary loss guides the feature recovery process of the local decoding branch, PALA is used to optimize the high-dimensional hyperparameters of DLCE-Net to improve its reliability, and the MW-GCA algorithm is used to obtain daylily picking information. The dual-branch decoder includes a 3C local feature decoding module, an upsampling module and an auxiliary loss; the PCEN includes a residual structure, a multi-scale pooling category information extraction module and an upsampling module; the PLFM includes strip pooling, dilated convolutions with different dilation rates and a CBR module; the MW-GCA includes a Mini-Window local window mechanism, a "U"-shaped feature window screening algorithm based on VASU and a cropping angle calculation method.

2. The method for classifying pixels of densely planted daylilies and obtaining harvesting information according to claim 1, characterized in that: Construct the semantic segmentation model of ADLCE-Net dual-branch decoding as follows: In the daylily picking scenario, to address the feature extraction confusion problem caused by dense plant occlusion, the present invention constructs a five-stage backbone structure: the first four stages are constructed by stacking a downsampling module and an Efficient ViM module. After scaling the image size through convolution regularization, the Efficient ViM module gradually mines multi-scale deep semantic features, providing rich information for dual-branch decoding; the fifth stage is constructed by a bilinear interpolation upsampling module and an Efficient ViM module, which restores the feature map size and extracts high-level global semantics, supports PCEN category information extraction, and finally completes the feature extraction process of the backbone network; Then, in terms of the decoder, the present invention constructs a dual-branch decoder that complements local and global information. In the local detail decoding branch, the 3C module is designed to achieve effective fusion of multi-scale features of daylily, and then complete the reconstruction of occluded edge details, ensuring the complete recovery of fine structures within the class. The global context branch cleverly integrates the EfficientViM Block module carried by the encoder. First, it restores the spatial size of the feature map through upsampling operations, and then uses 1×1 convolution to complete cross-channel information interaction. Finally, it is fused with the feature map output by the local detail branch to construct a global context representation that includes the overall structural relationship and relative position information of the plant, thereby forming a hierarchical and rich multi-dimensional feature representation; Finally, a weighted auxiliary loss (AUX Loss) was designed to provide semantic supervision for the feature recovery process in the local detail decoding branch, promoting the refined recovery of local detail features. Furthermore, a PALA optimization strategy was designed to adaptively search for hyperparameters of the DLCE-Net model. By dynamically adjusting key parameters during network training, a self-optimizing ADLCE-Net segmentation model was developed. This strategy effectively enhanced the model's adaptability to the complex semantic structure of daylily images, further improving its reliability and robustness in segmentation tasks.

3. The method for classifying pixels and obtaining harvesting information of densely populated daylilies in a field according to claim 2, characterized in that: Construct a two-branch decoder as follows: In response to the problems of low feature utilization efficiency and significant semantic information loss in single-branch decoders in dense occlusion scenarios, the present invention constructs a dual-branch decoding structure with complementary local details and global context information. In terms of the local detail branch, the encoding result of the feature map will first enter the 3C (Cat+Channel Attention+CBR) module, which achieves pixel-level detail refinement by fusing the feature information of daylily images of different scales, where CBR represents Conv+BatchNorm+ReLU. The specific way in which the 3C module fuses the feature information of daylilies of different scales is as follows: first, feature alignment and splicing operations are performed; second, Channel Attention screening is performed on the fused feature map to remove the feature information of useless channels; finally, the number of channels of the feature map is adjusted through the CBR module to restore it to the number of channels before input. The mathematical expression of its specific calculation is as follows: ； ； ； ； ; ; in, and Respectively represent the output results of the encoder of the current stage and the stage after the current stage; Indicates that the feature map and Splice by channel; and represents a fully connected layer with shared parameters; Represents the Sigmoid activation function; express Convolution kernel; In terms of global context branch, the present invention integrates the EfficientViM Block module, which is obtained by upsampling operation and 1x1 convolution with the local detail branch. By fusion, we can achieve global distribution perception of different feature images of daylily, and thus realize global feature recovery. The specific mathematical expression is as follows: ; ; Among them, when hour, Represents the output result of the previous global decoding stage, which ; Especially, when hour, After decoding in the EfficientViM Block module, the global context decoding branch The output of the decoding stage The specific mathematical expression of the calculation is shown above. express Convolution kernel. Finally, the final output of the global context decoder is The daylily plant category information extracted by the PCEN module Perform adaptive fusion to generate the final semantic segmentation result. The specific mathematical expression is as follows, where: represents the learnable adaptive fusion parameters; ; 。 4. The method for classifying pixels and obtaining harvesting information of densely populated daylilies in a field according to claim 2, characterized in that: Construct the PCEN multi-scale parallel category focusing module as follows: To address the problem of blurred occlusion segmentation boundaries and misjudgment of labels caused by discrete features within categories and similar features between categories in the differentiation of maturity of daylily, the present invention proposes a multi-scale hierarchical category enhancement module (PCEN). By multi-scale pooling and hierarchical sampling of high-level semantic features, a hierarchical feature aggregation structure is constructed to extract key discriminant features, suppress redundancy and strengthen category difference discrimination, thereby improving the segmentation performance of densely occluded scenes. For the input high-level features, they are first input into the multi-scale pooling module, and three pooling kernels with kernel sizes of 3×3, 7×7 and 11×11 are set for feature extraction. For each pooling kernel, maximum pooling (MaxPool) and average pooling (AvgPool) operations are performed respectively, and the advantages of the two pooling methods are combined to obtain a richer multi-scale feature representation. Subsequently, the extracted multi-scale features are subjected to residual fusion to retain image detail information and contextual semantic association. Finally, after a 2x upsampling operation and 1×1 convolution processing, the spatial alignment of high-level semantic information and the output features of the dual-branch decoder is achieved, providing standardized input for the subsequent feature fusion with the segmentation head. The output of the module PCEN The mathematical expression of the specific calculation is as follows: ; ; ; in Indicates a 2x upsampling operation, and The corresponding kernel sizes are Max pooling and average pooling.

5. The method for classifying pixels and obtaining harvesting information of densely planted daylilies in a field according to claim 2, characterized in that: Construct a PLFM multi-resolution multi-receptive field progressive step-by-step feature mining module as follows: To address the difficulty in defining boundaries and resulting in large segmentation errors caused by overlapping occlusions of daylily plants in densely planted scenes, this paper proposes a progressive ladder-like feature mining module (PLFM). Through multi-receptive field and multi-resolution progressive feature mining, this module extracts key features that are often overlooked in dense occlusion and achieves refined deep-level feature extraction, enhancing the model's perception of dense scenes. Specifically, the PLFM module consists of three similar feature extraction submodules. Each submodule first processes input features using dilated convolutions with different dilation rates. Because convolutions with different dilation rates have different receptive fields, they can extract feature information at different resolutions in dense scales. Next, a combined pooling framework is constructed to sample the processed features, further extracting important information and reducing the interference of redundant information on model training. During the pooling process, this paper constructs a combined pooling framework that integrates max pooling (MaxPool), average pooling (AvgPool), and strip pooling (StripPool). Through the collaborative design of these three pooling operations, it effectively aggregates multi-dimensional feature information. Among them, strip pooling can effectively capture long-range dependencies while also handling strip-shaped targets, making it very suitable for the daylily segmentation task. By stacking the three submodules with a residual connection structure, the PLFM module can effectively enhance the model's ability to perceive the characteristics of daylily plants in densely intertwined and mutually occluded scenes, thereby improving segmentation accuracy. For the input of a certain stage of the encoder, , the output after processing by the PLFM module The mathematical expression of the specific calculation is as follows: ; ; ; ; ; ; ; in, represents the dilated convolution with a dilation rate of r, refers to the ReLU activation function, and Represent maximum pooling and average pooling respectively, Stands for strip pooling.

6. The method for classifying pixels and obtaining harvesting information of densely populated daylilies in a field according to claim 2, characterized in that: Design the main and auxiliary collaborative loss strategy as follows: The specific mathematical expression of the loss function designed by the present invention is as follows: ; In the local detail decoding branch, the auxiliary loss function is used to apply a refined semantic supervision signal to guide the feature refinement recovery process of the branch, effectively enhancing the semantic consistency of multi-scale features, thereby improving the overall segmentation accuracy of the complex structure of the daylily plant. In addition, in order to clarify the leading role of the main loss function, the present invention designs a weight parameter It is used to adjust the weight of the auxiliary loss term in the total loss function. This parameter will be determined in the subsequent DLCE-Net hyperparameter self-optimization process. and The specific mathematical expression of the calculation is as follows: ; ; The auxiliary loss function is the same as the main loss function of DLCE-Net, which is composed of the cross entropy loss function and Loss Function Composition. and The specific formula is as follows, where K represents the total number of categories predicted by the segmentation model. represents the number of samples, Indicates in Among the samples, the The true label of the class, Indicates in Among the samples, the model predicts that The probability value of the class, ; ; ; The input of the auxiliary loss function is the decoding result of the last stage in the local detail decoding structure ,exist Before inputting into the auxiliary loss function, it needs to be processed. First, a 1x1 convolution layer is used to adjust the number of channels of the image to the total number of categories predicted by the segmentation model. ,Then, the image is restored to the input resolution through upsampling. The specific mathematical expression of the calculation is as follows, where, Represents the input value of the auxiliary loss function; 。 7. The method for classifying pixels and obtaining harvesting information of densely planted daylilies in a field according to claim 2, characterized in that: Construct the PALA hyperparameter optimization strategy as follows: The PALA strategy proposed in this invention deeply condenses the core idea of ​​the ALA algorithm, and conducts three-dimensional innovative optimization to address the problems of slow convergence, easy to fall into local optimality, weak adaptability of complex constraints, and global / local search imbalance in the hyperparameter self-optimization task of the DLCE-Net model: a population initialization mechanism based on a good point set is designed, and through mathematical space measurement and coordinate mapping, the initial population is made to have a measure-balanced distribution in high-dimensional space, laying a solid foundation for global exploration; a parallel artificial lemming optimization algorithm system is constructed, the population is deconstructed into three types of differentiated sub-populations, spatial decoupling is achieved through parameter configuration, and the optimal solution is directed across the population with the help of a circular neighborhood interaction mechanism, thereby improving the convergence performance of high-dimensional non-convex space; an elite inversion system is built, and elite individuals are extracted through percentile screening to generate a mirror solution space, and new search vectors are injected through a two-way preferential replacement strategy to suppress premature convergence. The three-dimensional innovative collaborative construction of a "balanced initialization-collaborative search-inversion evolution" closed-loop optimization system significantly improves the efficiency of hyperparameter optimization under complex constraints and the generalization ability of the model; In the construction of the DLCE-Net model, the present invention establishes DLCE-Net as the core architecture and proposes the PALA algorithm as a hyperparameter optimization strategy to achieve systematic tuning of the model hyperparameters. The optimization process selects 14 key hyperparameters as decision variables of the PALA algorithm, specifically including: optimizer category (sgd, adam), initial learning rate size (lr), momentum coefficient (momentum), learning rate reduction method (step, cos), auxiliary loss weight coefficient (λ), and 9 expansion rates (Expand_ratio) of the HSM-SSD module in the encoder and decoder. Under the PALA algorithm framework, each hyperparameter combination corresponds to an individual in the algorithm, and all candidate combinations constitute the initial population. Parameter optimization is achieved through the algorithm's adaptive search mechanism, and finally the ADLCE-Net model is constructed. The fitness function serves as a guide for algorithm search and is used to quantitatively evaluate individual performance. The fitness function defined for the DLCE-Net hyperparameter optimization task has the following mathematical expression: ； ； in, The function indicates that the model training is obtained under the specified 14-dimensional hyperparameter setting. value, Represents the fitness function. The individual with the best fitness will be used as the hyperparameter setting of ADLCE-Net.

8. The method for classifying pixels and obtaining harvesting information of densely populated daylilies in a field according to claim 2, characterized in that: The MW-GCA daylily picking point intelligent positioning algorithm is designed as follows: In high-density planting of daylilies, the plant posture and terminal morphology are complex, and there are leaf occlusions, background interference, and stem bending, and the flower morphology and color are variable, which makes it difficult for traditional picking point positioning algorithms to ensure robustness. To address this problem, the present invention proposes the MW-GCA daylily picking point intelligent positioning algorithm. First, extract the mask merge image of the mature and immature areas , and then obtain The geometric coordinates of the corner points of the image, the specific mathematical expression is as follows. In order to solve the problem of reduced positioning accuracy due to insufficient integrity of the edge area mask, the boundary corner points are optimized. Then, the potential picking points are located through the independently constructed corner point detection strategy, and the corner point filtering strategy is implemented for each connected area, focusing on eliminating invalid corner points at the top of the plant, and focusing the analysis on the feature-significant area at the bottom of the plant. The local feature analysis window is constructed through the innovative Mini-Window mechanism, which can effectively block irrelevant background interference outside the window, allowing the algorithm to focus on the key features of the target area; ； Specifically, the Mini-Window refers to a tiny local window created near a corner point, used to analyze features near the corner point and select corner points that meet the characteristics of the picking point. In this algorithm, the present invention selected a 60-pixel square as the side length of the Mini-Window. After in-depth research on the corner points at the picking point, it was found that the corner points that meet the picking point requirements all have two common characteristics: First, if the Mini-Window is separated from the vertical symmetry axis, the white area is mostly concentrated in the upper half, while the lower half is the black area; second, the upper half of the Mini-Window exhibits a "U"-shaped feature, and this "U"-shaped area exists in three situations: tilted to the left, tilted to the right, and centered. Based on the above-mentioned neighborhood feature constraints, the local window of the candidate corner point is screened for geometric and grayscale distribution features to ultimately determine the coordinates of the picking point of the daylily plant.

9. The method for classifying pixels and obtaining harvesting information of densely populated daylilies in a field according to claim 2, characterized in that: Design a method for calculating the cutting angle of daylily based on Mini-Window as follows: In the picking of daylilies, the selection of cutting angle is the key to the picking quality. A reasonable angle can reduce plant damage, maintain the appearance, and thus improve the economic benefits of automated picking. To this end, the present invention constructs a cutting angle calculation method based on Mini-Window. First, through the Mini-Window at the picking point, the two intersection points where the white block intersects with the Mini-Window boundary are extracted. 、 ,remember for and Then, connect the midpoints of and , which is the posture line segment of the daylily plant corresponding to the picking point, recorded as Finally, draw a line perpendicular to The line segment is recorded as By calculation The angle between the vertical right line segment and the vertical right line segment can determine the final cutting angle of the picking point. The specific mathematical expression of the calculation is as follows: ； ； ； ； ； ； in, and The function is used to obtain the left and right intersection points of the white block and the Mini_Window boundary. Function used to draw connections and A straight line, The function is used to draw a perpendicular line segment of a line segment. The function is used to calculate the angle between a line segment and a horizontal line segment to the right.

10. The method for classifying pixels and obtaining harvesting information of densely planted daylilies in a field according to claim 2, characterized in that: The model backbone structure Backbone accepts a 512×512 daylily image.

Images