Winter melon seedling salt tolerance phenotype grading method based on deep visual perception network

Abstract

The application discloses a winter melon seedling salt tolerance phenotype grading method based on a deep visual perception network, and belongs to the technical field of cross of wisdom agriculture and computer vision. The technical scheme is as follows: a winter melon seedling image dataset containing four salt tolerance grades is constructed; a deep visual perception network GYSalt-Net is constructed and trained, the network simultaneously captures macro wilting and micro lesion features by integrating a global-local hybrid feature extraction module, introduces a partial convolution adaptive context enhancement module to suppress complex background noise and enhance leaf region features, and adopts a dynamic up-sampling module based on point sampling to realize lightweight of the model and high-fidelity reconstruction of features; and the trained network is used to infer the salt tolerance grade of the seedling image. The application realizes high-throughput, automatic and accurate grading of the winter melon seedling salt tolerance, and the model is lightweight and suitable for deployment of portable computing devices.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of smart agriculture and computer vision, and in particular relates to a method for grading the salt tolerance phenotype of winter melon seedlings based on deep visual perception networks. Background Technology

[0002] Soil salinization is one of the major abiotic stressors limiting global agricultural productivity and posing a serious threat to food security. As an important vegetable crop, winter melon's development is also constrained by salinization. Breeding salt-tolerant varieties is the core approach to solving this problem, and accurate and efficient identification of salt tolerance phenotypes is a crucial foundation for successful breeding.

[0003] Currently, methods for identifying salt tolerance in winter melon have mainly gone through three stages of development, but all of them have significant technical bottlenecks. Traditional methods rely on manually measuring morphological indicators such as plant height and root length, or determining physiological and biochemical indicators such as antioxidant enzyme activity. These methods usually require destructive sampling, cannot continuously track dynamic changes in the same plant, and are cumbersome and have low throughput, making it difficult to meet the efficiency requirements of modern large-scale breeding. Manual visual grading is greatly affected by subjective factors and is difficult to quantify subtle phenotypic differences such as leaf yellowing and curling.

[0004] With technological advancements, high-throughput phenotypic analysis based on RGB imaging or chlorophyll fluorescence imaging has been introduced. However, these single-modality imaging analysis methods suffer from limitations due to their limited dimensionality. For instance, ordinary RGB image analysis struggles to effectively distinguish between color changes caused by normal plant growth and development and pathological damage resulting from salt stress; while a single chlorophyll fluorescence parameter is insufficient to reveal the physiological responses in core energy conversion processes such as thylakoid membrane proton dynamics during the early stages of salt stress. Furthermore, the relevant specialized imaging equipment is typically expensive, complex to operate, and time-consuming, hindering its widespread, low-cost application in field settings.

[0005] In recent years, general-purpose deep learning object detection models, represented by the YOLO series, have begun to be applied to crop phenotypic analysis. However, when these general-purpose models are directly transferred to the specific task of detecting salt tolerance in winter melon, they face a significant challenge of "mismatch between visual perception and physiological characteristics." First, in the early stages of salt stress, plants exhibit extremely subtle visual features such as sub-pixel micro-curling of leaf margins and high-frequency yellowing patterns on the leaf surface. The backbone network of general-purpose models lacks an effective extraction mechanism for such fine-grained features, easily leading to missed detections. Second, in actual breeding scenarios, image backgrounds are complex, containing soil, mulch, weeds, and densely packed seedlings. The feature fusion module of general-purpose models struggles to accurately focus on leaf lesion areas against such complex backgrounds, resulting in a high false detection rate. Finally, complex models designed for high accuracy often have a large number of parameters and computational redundancy, making them difficult to adapt to edge computing devices with limited computing power, such as agricultural inspection robots or handheld terminals, thus limiting their real-time, low-cost deployment and application in the field.

[0006] Therefore, there is an urgent need for a method that can deeply understand the physiological mechanism of salt damage and utilize advanced vision technology to achieve high-throughput, lightweight, and high-precision intelligent detection and grading of salt tolerance in winter melon seedlings in complex field environments. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a phenotypic grading method for salt tolerance in winter melon seedlings based on a deep visual perception network, thereby resolving the issues present in the prior art.

[0008] In a first aspect, to achieve the above objectives, the present invention provides a method for grading the salt tolerance phenotype of winter melon seedlings based on a deep visual perception network, comprising the following steps: Construct a dataset containing growth images of winter melon seedlings with different salt tolerance levels; A deep visual perception network was constructed and trained. The deep visual perception network integrates a global-local hybrid feature extraction module for simultaneously extracting macroscopic wilting and microscopic lesion features, a partially convolutional adaptive context enhancement module for enhancing leaf region features and suppressing background noise in complex backgrounds, and a dynamic upsampling module for achieving lightweight high-fidelity feature reconstruction. Using a trained deep visual perception network, inference is performed on the input image of winter melon seedlings to output their salt tolerance level.

[0009] Optionally, constructing the dataset includes: collecting multiple days of growth images of winter melon seedlings after being subjected to salt stress during their growth period, labeling the plant areas in the images, and classifying them into four levels based on their salt tolerance: normal, highly tolerant, moderately tolerant, and sensitive, thus forming a training set, a validation set, and a test set.

[0010] Optionally, the global-local hybrid feature extraction module works by: performing downsampling and depthwise convolution preprocessing on the input features, then performing global context feature extraction and local texture feature extraction in parallel, and finally fusing the two types of extracted features and outputting them through residual connections.

[0011] Optionally, the global context feature extraction is achieved using adaptive average pooling along the X and Y directions; the local texture feature extraction is achieved using depthwise separable convolution combined with average pooling.

[0012] Optionally, the working process of the partially convolutional adaptive context enhancement module includes: performing partial convolution operations on the input feature map to reduce computational cost and preserve spatial information; and jointly enhancing the weight of leaf region features by establishing context enhancement paths that establish dependencies between features and fully connected layer paths that adaptively enhance discriminative features.

[0013] Optionally, the partial convolution operation is performed only on a portion of the channels of the input feature map, while the remaining channels remain unchanged.

[0014] Optionally, the dynamic upsampling module adopts a point-based sampling mechanism, which dynamically generates the positions of upsampling points based on the semantic content of the input features to reconstruct the feature map.

[0015] Optionally, the deep visual perception network is reconstructed based on a single-stage object detection architecture, and the global-local hybrid feature extraction module, the partially convolutional adaptive context enhancement module, and the dynamic upsampling module are integrated into its backbone network, neck network, and upsampling part, respectively.

[0016] Secondly, the present invention also provides a computer terminal device, comprising: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the method for grading the salt tolerance phenotype of winter melon seedlings based on a deep visual perception network in the first aspect described above.

[0017] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the method for grading the salt tolerance phenotype of winter melon seedlings based on a deep visual perception network in the first aspect described above.

[0018] Compared with the prior art, the present invention has the following advantages and technical effects: This invention provides a phenotypic grading method for salt tolerance in winter melon seedlings based on deep visual perception networks. By integrating a global-local hybrid feature extraction module, it effectively enhances the capture ability of fine yellowing textures and edge curling features of leaves in the early stages of salt stress, solving the problem of missed detection of small lesions. By introducing a partially convolutional adaptive context enhancement module, it significantly suppresses interference from complex backgrounds such as soil and weeds, improving the model's robustness and accuracy in densely planted environments. By applying a point-sampling-based dynamic upsampling module, it achieves high-fidelity reconstruction of key morphological features such as leaf margins while significantly reducing the overall number of model parameters and computational complexity, making the lightweight model adaptable and deployable on agricultural edge computing devices. Ultimately, this invention achieves high-throughput, automated, and accurate grading of salt tolerance in winter melon seedlings and provides a feasible technical path for building smart agricultural detection terminals based on low-cost hardware or recycled electronic devices. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the salt tolerance dataset of winter melon seedlings according to an embodiment of the present invention; Figure 2 This is a statistical chart of the label distribution of the winter melon salt tolerance dataset according to an embodiment of the present invention, wherein (a) is a distribution chart of the number of categories, (b) is a distribution chart of the bounding box shape, (c) is a distribution chart of the target center point location, (d) is a distribution chart of the target size, and (e) is a label correlation chart; Figure 3 This is a schematic diagram of the GYSalt-Net deep vision perception model structure according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the GL-HFE module structure according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the PACE module structure according to an embodiment of the present invention; Figure 6 This is a schematic diagram comparing the PR curves of the model before and after the improvement in this embodiment of the invention, where (a) is the original YOLOv11n and (b) is the mAP@0.5 of GYSalt-Net; Figure 7 This is a comparison diagram of the confusion matrix of an embodiment of the present invention, where (a) is the original model and (b) is the GYSalt-Net model; Figure 8 This is a schematic diagram of the F1-Confidence curve in an embodiment of the present invention; Figure 9This diagram illustrates how the optimal F1 score of GYSalt-Net in this embodiment of the invention is improved from 0.81 to 0.87 compared to the baseline model. Figure 10 This is a schematic diagram illustrating the accuracy of bounding box detection in the YOLOv11n model according to an embodiment of the present invention. Figure 11 This is a schematic diagram of the recall rate of bounding box detection in the YOLOv11n model according to an embodiment of the present invention; Figure 12 This is a schematic diagram of the average accuracy of the YOLOv11n model in this embodiment of the invention when the IoU threshold is 0.5; Figure 13 This is a schematic diagram illustrating the average accuracy of the YOLOv11n model with IoU thresholds ranging from 0.5 to 0.95 in an embodiment of the present invention. Figure 14 This is a schematic diagram illustrating the accuracy of bounding box detection in the GYSalt-Net model according to an embodiment of the present invention. Figure 15 This is a schematic diagram of the recall rate of bounding box detection in the GYSalt-Net model according to an embodiment of the present invention; Figure 16 This is a schematic diagram of the average accuracy of the GYSalt-Net model in this embodiment of the invention when the IoU threshold is 0.5; Figure 17 This is a schematic diagram illustrating the average accuracy of the GYSalt-Net model with IoU thresholds ranging from 0.5 to 0.95 in an embodiment of the present invention. Figure 18 A visual comparison diagram of the actual labeled detection results; Figure 19 A visual comparison diagram of YOLOv11n prediction results; Figure 20 A visual comparison diagram of the prediction results from GYSalt-Net. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] Example 1 This embodiment provides a method for grading the salt tolerance phenotype of winter melon seedlings based on a deep visual perception network, including: constructing a dataset containing growth images of winter melon seedlings with different salt tolerance levels; A deep visual perception network was constructed and trained. The deep visual perception network integrates a global-local hybrid feature extraction module for simultaneously extracting macroscopic wilting and microscopic lesion features, a partially convolutional adaptive context enhancement module for enhancing leaf region features and suppressing background noise in complex backgrounds, and a dynamic upsampling module for achieving lightweight high-fidelity feature reconstruction. Using a trained deep visual perception network, inference is performed on the input image of winter melon seedlings to output their salt tolerance level.

[0023] Furthermore, constructing the dataset includes: collecting multiple days of growth images of winter melon seedlings after being subjected to salt stress during their growth period, labeling the plant areas in the images, and classifying them into four levels based on their salt tolerance: normal, highly tolerant, moderately tolerant, and sensitive, thus forming a training set, a validation set, and a test set.

[0024] Specifically, the implementation process of this embodiment includes: Step S1: Construct a dataset of salt tolerance phenotypes for winter melon. Images of winter melon seedlings were collected for nine consecutive days after they reached the one-leaf-one-heart stage (simulating salt stress after transplanting). The sample included both the salt stress treatment group and the no-salt stress control group. The acquired images were professionally labeled to accurately mark the regions of the winter melon plants, and they were divided into four levels according to salt tolerance: Normal, Highly Salt-tolerant, Moderately Salt-tolerant, and Salt-sensitive. Based on this, a standardized dataset containing training, validation, and test sets was constructed.

[0025] Furthermore, the working process of the global-local hybrid feature extraction module includes: after downsampling and depthwise convolution preprocessing of the input features, global context feature extraction and local texture feature extraction are performed in parallel, and then the two types of extracted features are fused and output through residual connection.

[0026] Furthermore, the global context feature extraction is achieved using adaptive average pooling along the X and Y directions; the local texture feature extraction is achieved using depthwise separable convolution combined with average pooling.

[0027] Specifically, the implementation process of this embodiment includes: Step S2: Construct the GYSalt-Net deep visual perception model ( Figure 3 ): The GYSalt-Net constructed in this invention is not a traditional convolutional neural network, but a deep visual perception network that simulates the visual cognitive process of plant phenotyping experts. This deep visual perception network is reconstructed based on the YOLOv11n architecture and integrates three core biomimetic visual modules: S2-1, Backbone Network Improvement – ​​Introducing, for example Figure 4 The GL-HFE module shown (Global-Local Hybrid Visual Feature Extraction) is as follows: For the phenotype of damage to winter melon, the GL-HFE module simulates the observation logic of human experts, which combines "panoramic overview" with "detailed focus".

[0028] A dual-branch parallel structure is adopted: ① Input processing: Downsampling is performed using convolution with a stride of 2, followed by feature preprocessing using 5×5 depthwise convolution (DepthwiseConv), which reduces the number of parameters while expanding the receptive field.

[0029] ② Global Branch: Utilizes adaptive average pooling in the X and Y directions to establish a "global visual context," capturing spatial dependencies in the horizontal and vertical directions, and focusing on modeling the overall macroscopic geometric morphology of leaves, such as wilting and lodging.

[0030] ③Local Branch: Utilizes depthwise separable convolution combined with average pooling to focus on extracting "local visual textures", such as leaf vein chlorosis, necrotic spots, and edge curling features.

[0031] ④ Fusion Output: The first branch extracts global contextual features through spatial adaptive average pooling; the second branch extracts local texture features through depthwise separable convolution; the outputs of the two branches are concatenated and then fused with the input through a residual connection. After feature fusion, the output through a residual connection achieves a complete visual representation of the image semantics.

[0032] Furthermore, the working process of the partially convolutional adaptive context enhancement module includes: performing partial convolution operations on the input feature map to reduce computational cost and preserve spatial information; and jointly enhancing the weight of leaf region features by establishing context enhancement paths that establish dependencies between features and fully connected layer paths that adaptively enhance discriminative features.

[0033] Furthermore, the partial convolution operation process involves performing convolution operations only on a portion of the input feature map channels, while leaving the remaining channels unchanged.

[0034] Specifically, the implementation process of this embodiment includes: S2-2, Neck Network Improvement – ​​Introducing, for example Figure 5The PACE module shown (partial convolutional adaptive visual attention enhancement): To address visual interference from soil and weeds, the PACE module constructs a "visual attention filtering mechanism" that simulates the "selective attention" of human vision.

[0035] Triple mechanism: ① Partial Convolution (PConv): Convolution operations are performed only on 1 / 4 channels of the input feature map, while the remaining channels remain unchanged. This aims to reduce computational redundancy while preserving the original spatial information using the unprocessed channels, thus initially filtering out soil and weed background.

[0036] ② Context Enhancement: A dual-path parallel architecture is adopted. The main path uses ContextBlock to establish dependencies between features; the secondary path adaptively enhances discriminative features that are strongly correlated with salt tolerance through fully connected layers (FC).

[0037] ③ Visual focus guidance: The above mechanism guides the network to focus the "visual focus" on the leaf area, suppresses the response weights of background noise, and improves the detection robustness in complex scenes.

[0038] Furthermore, the dynamic upsampling module adopts a point-sampling-based mechanism, which dynamically generates the positions of upsampling points based on the semantic content of the input features to reconstruct the feature map.

[0039] Specifically, the implementation process of this embodiment includes: S2-3, Upsampling Optimization – Introducing the DySample Module (Dynamic Reconstruction Based on Visual Content Awareness): To address the image blurring caused by traditional upsampling, the DySample module employs a "content-aware" sampling strategy.

[0040] Instead of using fixed convolutional kernels, this method employs a point sampling mechanism that dynamically generates sampling points by learning the distribution of input features. It dynamically adjusts the sampling positions based on the semantic content of the feature map, much like how human vision automatically adjusts its focus when seeing an object clearly. This significantly reduces computational cost (GFLOPs) while achieving high-fidelity visual reconstruction of fine details such as curled leaf edges, avoiding jagged edges.

[0041] Furthermore, the deep visual perception network is reconstructed based on a single-stage object detection architecture, and integrates the global-local hybrid feature extraction module, the partially convolutional adaptive context enhancement module, and the dynamic upsampling module into its backbone network, neck network, and upsampling part, respectively.

[0042] Step S3, Model Training: Input the dataset constructed in step S1 into the GYSalt-Net model ( Figure 3 End-to-end training is performed using stochastic gradient descent (SGD) or the AdamW optimizer, iteratively updating the weights through the loss function until the model converges.

[0043] Step S4, Intelligent Hierarchy and Reasoning: The trained GYSalt-Net is used to perform inference on the images of winter melon seedlings to be tested. The model outputs the local bounding box of the plant and the salt tolerance level, realizing automated and high-throughput phenotypic analysis.

[0044] Example 2 In this embodiment, a computer terminal device is provided, including: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-described method for grading the salt tolerance phenotype of winter melon seedlings based on a deep visual perception network.

[0045] In this embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described method for grading the salt tolerance phenotype of winter melon seedlings based on a deep visual perception network.

[0046] Figure 1 This is an example of a dataset on the salt tolerance of winter melon seedlings. Figure 1 (a) Based on salt tolerance phenotype, seedlings were classified into four grades: Normal, Salt-tolerant, Salt-intolerant, and Salt-sensitive. Figure 1 (b) The leaf phenotype on day 3 (D3) of salt stress only showed faint yellowing patterns between the veins and very small initial chlorotic spots, which are precursor characteristics that are difficult to identify with the naked eye and conventional models. Figure 1 (c) The leaf phenotype of the same plant on day 9 (D9) shows significant necrotic patches and wilting symptoms, which are distinctive and easy to identify at this time.

[0047] Figure 2 This is a statistical chart showing the label distribution of the winter melon salt tolerance dataset. Figure 2 As can be seen from the (a) category distribution plot, the dataset covers four salt tolerance levels and exhibits a long-tailed distribution. Figure 2 As can be seen from the bounding box shape distribution diagram in (b), most of the boxes are vertically elongated rectangles, indicating that the shapes of the labeled targets are relatively consistent. Figure 2Target center point location distribution map (c) and Figure 2 The distribution of center points in the label correlation plot (e) reveals the double-row arrangement of plants in the seedling tray. Figure 2 The target size distribution map (d) shows that the aspect ratio of the labeled targets is linearly positively correlated, covering different size changes of seedlings from germination to the one-leaf-one-heart stage. Overall, the dataset has high labeling quality and clear distribution patterns, providing a solid data foundation for training a highly robust GYSalt-Net.

[0048] Figure 6 This is a diagram illustrating the comparison of PR curves before and after the model was improved. Figure 6 (a) is 84.2% of the original YOLOv11n, and (b) is 92.8% of the mAP@0.5 of GYSalt-Net, especially on the Salt-intolerant category, where the AP value surged from 0.706 to 0.874.

[0049] Figure 7 This is a comparison chart of confusion matrices, where Figure 7 (a) shows that the original model is highly prone to misclassifying 'medium-intolerant' as 'high-tolerant', with a misclassification rate as high as 37%. Figure 7 (b) shows that the GYSalt-Net model has significantly enhanced feature discrimination ability in complex backgrounds. The correct recognition rate of this category has increased dramatically from 47% to 75%, effectively solving the confusion problem caused by high phenotypic similarity.

[0050] Figure 8 This is a schematic diagram of the F1-Confidence curve. Figure 9 The diagram illustrates how GYSalt-Net's best F1 score improved from 0.81 to 0.87 compared to the baseline model. The worst-performing Salt-intolerant class (green line) in the baseline model was significantly improved in GYSalt-Net, with its curve showing a substantial improvement, demonstrating that GYSalt-Net effectively solves the inter-class confusion problem in complex contexts. Figure 10 This is a diagram illustrating the accuracy of bounding box detection in the YOLOv11n model. Figure 11 This is a schematic diagram illustrating the recall rate of bounding box detection in the YOLOv11n model. Figure 12 This is a schematic diagram illustrating the average accuracy of the YOLOv11n model when the IoU threshold is 0.5. Figure 13 This diagram illustrates the average accuracy of the YOLOv11n model with IoU thresholds ranging from 0.5 to 0.95. Figure 14 This diagram illustrates the accuracy of bounding box detection in the GYSalt-Net model (higher accuracy means more accurate detection). Figure 15This diagram illustrates the recall rate (higher values ​​indicate more complete detection) for bounding box detection in the GYSalt-Net model. Figure 16 The diagram shows the average accuracy of the GYSalt-Net model when the IoU threshold is 0.5. It can be seen that the mAP50 curve converges faster and the final value is higher (>92%), indicating that GYSalt-Net not only improves the detection limit but also speeds up the convergence efficiency of the model. Figure 17 This is a schematic diagram showing the average accuracy (reflecting how accurately the bounding box position is regressed) of the GYSalt-Net model with IoU thresholds ranging from 0.5 to 0.95. Figure 18 A visual comparison diagram of the actual labeled detection results; Figure 19 A visual comparison diagram of YOLOv11n prediction results; Figure 20 A visual comparison diagram of the prediction results from GYSalt-Net.

[0051] This invention provides a phenotypic grading method for salt tolerance in winter melon seedlings based on deep visual perception networks. By integrating a global-local hybrid feature extraction module, it effectively enhances the capture ability of fine yellowing textures and edge curling features of leaves in the early stages of salt stress, solving the problem of missed detection of small lesions. By introducing a partially convolutional adaptive context enhancement module, it significantly suppresses interference from complex backgrounds such as soil and weeds, improving the model's robustness and accuracy in densely planted environments. By applying a point-sampling-based dynamic upsampling module, it achieves high-fidelity reconstruction of key morphological features such as leaf margins while significantly reducing the overall number of model parameters and computational complexity, making the lightweight model adaptable and deployable on agricultural edge computing devices. Ultimately, this invention achieves high-throughput, automated, and accurate grading of salt tolerance in winter melon seedlings and provides a feasible technical path for building smart agricultural detection terminals based on low-cost hardware or recycled electronic devices.

[0052] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for grading the salt tolerance phenotype of winter melon seedlings based on deep visual perception networks, characterized in that, Includes the following steps: Construct a dataset containing growth images of winter melon seedlings with different salt tolerance levels; A deep visual perception network was constructed and trained. The deep visual perception network integrates a global-local hybrid feature extraction module for simultaneously extracting macroscopic wilting and microscopic lesion features, a partially convolutional adaptive context enhancement module for enhancing leaf region features and suppressing background noise in complex backgrounds, and a dynamic upsampling module for achieving lightweight high-fidelity feature reconstruction. Using a trained deep visual perception network, inference is performed on the input image of winter melon seedlings to output their salt tolerance level.

2. The method according to claim 1, characterized in that, The construction of the dataset includes: collecting growth images of winter melon seedlings after being subjected to salt stress treatment for several consecutive days, labeling the plant areas in the images, and classifying them into four levels based on their salt tolerance: normal, highly tolerant, moderately tolerant, and sensitive, thus forming a training set, a validation set, and a test set.

3. The method according to claim 1, characterized in that, The working process of the global-local hybrid feature extraction module includes: after downsampling and depthwise convolution preprocessing of the input features, global context feature extraction and local texture feature extraction are performed in parallel, and then the two types of extracted features are fused and output through residual connection.

4. The method according to claim 3, characterized in that, The global context feature extraction is achieved using adaptive average pooling along the X and Y directions; the local texture feature extraction is achieved using depthwise separable convolution combined with average pooling.

5. The method according to claim 1, characterized in that, The working process of the partially convolutional adaptive context enhancement module includes: performing partial convolution operations on the input feature map to reduce computation and preserve spatial information; and jointly increasing the weight of leaf region features by establishing context enhancement paths that establish dependencies between features and fully connected layer paths that adaptively enhance discriminative features.

6. The method according to claim 5, characterized in that, The partial convolution operation involves performing convolution operations only on a portion of the input feature map channels, while leaving the remaining channels unchanged.

7. The method according to claim 1, characterized in that, The dynamic upsampling module employs a point-sampling-based mechanism, which dynamically generates the positions of upsampling points based on the semantic content of the input features to reconstruct the feature map.

8. The method according to claim 1, characterized in that, The deep visual perception network is reconstructed based on a single-stage object detection architecture, and integrates the global-local hybrid feature extraction module, the partially convolutional adaptive context enhancement module, and the dynamic upsampling module in its backbone network, neck network, and upsampling part, respectively.

9. A computer terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the steps of the method as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-8.

Images