DSS-YOLO tomato flower pistil detection model and automatic pollination system

Abstract

The invention discloses a DSS-YOLO tomato flower pistil detection model and an automatic pollination system, and the model comprises a lightweight backbone network which is used for receiving original image data of a greenhouse tomato plant, carrying out the multi-scale fine-grained feature extraction of an image through a low-complexity convolution structure, and outputting an initial feature map; the efficient down-sampling module is used for carrying out dimension reduction and feature optimization on the initial feature map output by the lightweight backbone network and outputting an optimized high-discrimination-force feature map; and the detection head is connected to the output end of the high-efficiency down-sampling module, and adopts parallel classification branches and regression branches to respectively output the category probability and bounding box coordinates of the tomato flower stamens so as to complete a target detection task. By adopting the DSS-YOLO tomato flower pistil detection model and the automatic pollination system, the optimal balance among precision, speed and model lightweight is considered, and precise detection and efficient automatic pollination of tomato flower pistil in a greenhouse scene are realized.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to the DSS-YOLO tomato flower stamen detection model and automated pollination system. Background Technology

[0002] Currently, deep learning-based object detection technology has been widely used in agricultural plant phenotypic information acquisition, especially in the field of flower detection, where the YOLO series of algorithms has become the mainstream choice due to its balance between speed and accuracy. However, when specifically applied to the scenario of greenhouse tomato pollination robots, existing technologies have the following obvious defects and shortcomings: The contradiction between model complexity and deployment requirements: Existing mainstream detection models (such as YOLOv5, YOLOv8, etc.) are generally complex in structure, have a large number of parameters, and high computational overhead, making it difficult to achieve real-time inference on embedded devices of pollination robots with limited computing power, storage and power consumption, which seriously hinders their practical application.

[0003] Insufficient detection capability for small-scale and slender targets: Tomato flower stamens are the direct target of pollination, but they are extremely small in size, slender in shape, and have variable orientation. The feature extraction network and downsampling module of existing models are prone to losing detailed information about such targets. At the same time, common bounding box regression loss functions (such as CIoU) are not sensitive to orientation, resulting in poor flower stamen localization accuracy and high false negative rate.

[0004] Poor robustness in complex agricultural scenarios: Greenhouse environments present challenges such as dense foliage shading, overlapping flowers, variable lighting, and complex backgrounds. Existing models lack sufficient feature extraction and fusion capabilities, making it difficult to stably extract discriminative features in such complex scenarios, resulting in large fluctuations in detection performance and weak generalization ability. Summary of the Invention

[0005] The purpose of this invention is to provide a DSS-YOLO tomato flower and stamen detection model and an automated pollination system that achieves an optimal balance between accuracy, speed and lightweight model. It solves the problems of deployment difficulty, low accuracy, poor robustness and insufficient automation of pollination faced by existing technologies in greenhouse tomato flower and stamen detection, and realizes accurate detection and efficient automated pollination of tomato flowers and stamens in greenhouse settings.

[0006] To achieve the above objectives, this invention provides a DSS-YOLO tomato flower stamen detection model, comprising: a lightweight backbone network for receiving raw image data of greenhouse tomato plants, extracting multi-scale fine-grained features from the image through a low-complexity convolutional structure, and outputting an initial feature map; The high-efficiency downsampling module is embedded in the neck network of the DSS-YOLO tomato flower stamen detection model. It is used to reduce the dimensionality and optimize the features of the initial feature map output by the lightweight backbone network, and output the optimized feature map with high discriminative power. The detection head is used to process high-discrimination feature maps. It uses parallel classification and regression branches to output the category probability and bounding box coordinates of tomato flower stamens, respectively, to complete the target detection task.

[0007] Preferably, the lightweight backbone network is based on HGNetv2 architecture, and sequentially connects the initial feature extraction HGStem module, five stacked depthwise separable weighted gated (DWHGBlock) modules, the spatial pyramid pooling (SPPF) module, the channel spatial attention (C2PSA) module, and the depthwise separable convolution downsampling module.

[0008] Preferably, the DWHGBlock module includes five 3×3 depthwise separable convolutions, one 1×1 pointwise convolution, and an ESE attention mechanism layer connected in sequence.

[0009] Preferably, depthwise separable convolution includes channel-wise convolution and pointwise convolution, wherein channel-wise convolution is used for spatial feature extraction; pointwise convolution is used for channel dimension adjustment; the number of parameters for depthwise separable convolution is as follows: k×k×c+c×n; Where k is the kernel size; c is the number of input channels; and n is the number of output channels.

[0010] The preferred high-efficiency downsampling module is as follows: The input layer is used to receive the feature maps to be processed. Conv1: Contains Conv2d convolutional layers, BatchNorm2d batch normalization layers, and SiLU activation function, used to adjust the number of channels; Conv2: Contains Conv2d convolutional layers and BatchNorm2d batch normalization layers in sequence to further extract features; Output layer: The corresponding element values ​​of the feature maps of Conv1 and Conv2 are added together by element-wise addition to achieve feature reuse; the fused feature map is the final output of the module.

[0011] Preferably, the regression branch introduces a bounding box regression loss function, which includes a newly added angle penalty term. This angle penalty term guides the model to prioritize the regression direction by considering the vector angle between the ground truth box and the predicted box. The bounding box regression loss function is as follows: ; in, It is the bounding box regression loss function; It is the intersection-union ratio of the predicted bounding box and the ground truth bounding box; It is a shape penalty term, specifically expressed as: , It is a measure of width difference, specifically expressed as: , It is the width of the prediction box. It is the width of the actual frame; It is a measure of high variability, specifically expressed as: , It is the height of the predicted bounding box; It is the height of the predicted bounding box and the ground truth bounding box; It is a parameter that adjusts the degree of emphasis placed on shape loss; It is a distance penalty term, specifically represented as: , It is an intermediate variable in the distance metric. It is a horizontal distance metric. It is a vertical distance metric. It is the adjustment parameter for distance penalty. It is an overall distance metric. The normalized distance in the horizontal direction is specifically expressed as: , It is the x-coordinate of the center point of the prediction box. It is the width of the smallest bounding rectangle of the ground truth bounding box and the predicted bounding box. It is the x-coordinate of the center of the true bounding box. It is a real frame; It is the normalized distance in the vertical direction, specifically expressed as: , It is the y-coordinate of the center of the true bounding box. It is the ordinate of the center point of the prediction box. It is the height of the smallest bounding rectangle between the ground truth bounding box and the predicted bounding box.

[0012] Preferably, the distance penalty term Δ is obtained based on the angle penalty term, wherein the angle penalty term is as follows: ; in, Angle penalty item; This represents the height difference between the center points of the ground truth bounding box and the predicted bounding box. This represents the distance between the center points of the ground truth bounding box and the predicted bounding box.

[0013] The automated pollination system includes an image acquisition unit: used to acquire real-time image data of tomato plants in a greenhouse environment; Data preprocessing unit: used to process real-time image data and perform preliminary feature extraction to obtain model input data; Visual perception unit: integrates the DSS-YOLO tomato flower and stamen detection model, used to identify, classify and locate tomato flowers and stamens in the model input data, and output target category and bounding box coordinates; Pollination execution unit: Based on the bounding box coordinates output by the visual perception unit, it controls the execution structure to complete the precise pollination operation.

[0014] Therefore, the present invention employs the above-mentioned DSS-YOLO tomato flower stamen detection model, which has the following beneficial effects: (1) Through the collaborative design of a lightweight backbone network and an efficient downsampling module, the model achieves extreme lightweighting while maintaining high accuracy. Experiments show that compared with the benchmark model YOLOv11n, the weights, parameters, and computational cost of this model are reduced by 34%, 36%, and 35%, respectively, and the inference speed is increased to 65 frames / second, perfectly meeting the stringent deployment requirements of embedded devices.

[0015] (2) Through the feature fidelity of the efficient downsampling module and the angle-aware mechanism of the SIoU loss function, the present invention significantly enhances the model's ability to detect small-scale, slender targets such as tomato flower stamens. Test results on public datasets show that the model improves the average detection accuracy (mAP@0.5) and precision of flowers and stamens by 1% and 1.2% respectively, effectively reducing false negatives and false positives.

[0016] (3) The multi-scale feature fusion and efficient feature extraction mechanism of this invention enables it to better cope with challenges such as shading, stacking and light changes in the greenhouse environment, and shows more stable detection performance and stronger generalization ability than existing models.

[0017] This invention employs the aforementioned automated pollination system, directly addressing the automated pollination needs in smart agriculture. The system is plug-and-play and adaptable to various scenarios, and has positive practical significance for reducing agricultural labor costs, improving production efficiency, and promoting the intelligent transformation of agriculture.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a structural diagram of the DSS-YOLO model in an embodiment of the present invention; Figure 2 This is a schematic diagram of the depth-separable convolution (DWConv) principle in an embodiment of the present invention; Figure 3 This is a module structure diagram of the depth-separable weighted gate DWHGBlock in this embodiment of the invention; Figure 4 This is a network structure diagram of the high-efficiency downsampling module SCDown in an embodiment of the present invention; Figure 5 This is a geometric schematic diagram of the SIoU loss function in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] The present invention provides a DSS-YOLO tomato flower and stamen detection model, comprising: a lightweight backbone network for receiving raw image data of greenhouse tomato plants, extracting multi-scale fine-grained features from the image through a low-complexity convolutional structure, and outputting an initial feature map; The high-efficiency downsampling module is embedded in the neck network of the DSS-YOLO tomato flower stamen detection model. It is used to reduce the dimensionality and optimize the features of the initial feature map output by the lightweight backbone network, and output the optimized feature map with high discriminative power. The detection head is used to process high-discrimination feature maps. It uses parallel classification and regression branches to output the category probability and bounding box coordinates of tomato flower stamens, respectively, to complete the target detection task.

[0022] Example like Figure 1 As shown, the DSS-YOLO tomato flower stamen detection model of the present invention includes: The lightweight backbone network DWHGNetV2 is used to receive raw image data of greenhouse tomato plants, extract multi-scale fine-grained features from the images through a low-complexity convolutional structure, and output an initial feature map.

[0023] HGNetV2 is a high-performance backbone network that was initially used for RT-DETR object detection. Later, due to its excellent feature extraction capabilities and efficiency, it was widely used to improve detection models such as YOLO.

[0024] To overcome the computational bottleneck faced by the baseline model when deployed on mobile devices, this embodiment constructs a lightweight backbone network, DWHGNetv2, to replace the original backbone structure of YOLOv11n. Because HGNetv2 has advantages in maintaining high-resolution feature representation and excellent performance, the lightweight backbone network is based on the HGNetv2 architecture, with the initial feature extraction module HGStem sequentially connected in series for preliminary feature extraction and downsampling of the input image. Five stacked depth-separable weighted gated DWHGBlock modules are used to progressively extract and deepen feature representations by stacking five of these modules. The Spatial Pyramid Pooling SPPF module is used to fuse feature information at different scales and accelerate processing. The Channel Spatial Attention (C2PSA) module is used to enhance the feature responses of important channels and spatial locations; Depth-separable convolutional downsampling module.

[0025] The DWHGBlock module consists of five sequentially connected 3×3 depthwise separable convolutions (DWConv), one 1×1 pointwise convolution, and an ESE attention mechanism layer, as shown in the structure below. Figure 3 As shown, the DWHGBlock module is an improved version of HGNetv2. By replacing some of the standard convolutions with depthwise separable convolutions (DWConv), it significantly reduces the computational cost and parameter count while maintaining high-resolution feature representation.

[0026] Depthwise separable convolution (DWConv) includes channel-wise convolution and pointwise convolution. Channel-wise convolution is used for spatial feature extraction, while pointwise convolution is used for channel dimension adjustment. The principle diagram of DWConv is shown below. Figure 2 As shown.

[0027] The formula for calculating the parameters of depthwise separable convolution (DWConv) is as follows: k×k×c+c×n; Where k is the kernel size; c is the number of input channels; and n is the number of output channels.

[0028] DWConv effectively avoids redundant calculations across channels by splitting a single standard convolution into two steps: channel-wise convolution and point-wise convolution. The number of parameters is much lower than that of traditional standard convolution (k×k×c×n), which greatly improves computational efficiency. At the same time, while ensuring feature extraction capabilities, it builds a powerful multi-scale feature extraction capability, enhancing the perception of fine-grained features.

[0029] To address the issues of detail loss and high computational cost associated with traditional downsampling operations in greenhouse tomato flower stamen detection, this embodiment replaces the standard convolutional downsampling layer in YOLOv11 with the efficient SCDown module.

[0030] The efficient downsampling module SCDown is connected to the output of the lightweight backbone network DWHGNetv2 and embedded in the model neck network. It is used to receive the initial feature map output by the lightweight backbone network. While reducing the dimensionality of the initial feature map to reduce the computational cost of the model, it also reduces the loss of feature details of small-scale targets such as tomato flower stamens and outputs optimized feature maps with high discriminative power.

[0031] The high-efficiency downsampling module SCDown has the following structure: Figure 4 As shown.

[0032] The input layer is used to receive the feature maps to be processed.

[0033] Conv1: Contains, in sequence, Conv2d convolutional layers, BatchNorm2d batch normalization layers, and SiLU activation function, used to adjust the number of channels.

[0034] Conv2: Contains Conv2d convolutional layers and BatchNorm2d batch normalization layers in sequence to further extract features.

[0035] Output layer: By using element-wise summation, the element values ​​at corresponding positions in the feature maps of Conv1 and Conv2 are added together to achieve feature reuse. The fused feature map serves as the final output of the module, and its spatial dimensions meet the downsampling requirements: assuming a downsampling factor of k, the height and width of the output feature map will be reduced to 1 / k of the input.

[0036] This module employs a decoupled design that first adjusts the channel dimension through pointwise convolution and then achieves spatial reduction through depthwise convolution. This not only significantly reduces the number of parameters but also effectively preserves the texture and positional details of small-scale flower stamens, thereby improving the model's ability to extract subtle features and its overall detection robustness. Furthermore, it helps alleviate the gradient vanishing problem in deep networks and retains more details of the input information.

[0037] The detection head, connected to the output of the high-efficiency downsampling module SCDown, uses parallel classification and regression branches to output the category probability and bounding box coordinates of the tomato flower stamens, respectively, to complete the target detection task.

[0038] YOLOv11 uses the CIoU loss function by default for bounding box regression. This function, based on IoU, optimizes the localization of the predicted box by introducing penalty terms for center point distance and aspect ratio. However, CIoU exhibits significant limitations when detecting small-scale targets such as tomato flower stamens: firstly, because the aspect ratio of such targets has a limited range of variation, the aspect ratio penalty term in CIoU is difficult to provide effective gradient guidance; secondly, even a slight deviation in the center point of a small target can lead to a large relative localization error, and CIoU is not sensitive enough to such deviations. These two factors together affect the regression accuracy of the model for small targets.

[0039] Therefore, in this embodiment, the regression branch of the detection head introduces a bounding box regression SIoU loss function. This bounding box regression SIoU loss function includes a newly added angle penalty term. The angle penalty term guides the model to prioritize the regression direction by considering the vector angle between the ground truth box and the predicted box. The bounding box regression SIoU loss function is as follows: ; in, It is the bounding box regression loss function; It is the intersection-union ratio of the predicted bounding box and the ground truth bounding box; It is a shape penalty item; It is a distance penalty item.

[0040] like Figure 5 As shown, the core logic algorithm of the SIoU loss function is as follows: (1) Introduce an angle penalty term, as shown below: ; in, Angle penalty item; The height difference between the center points of the ground truth bounding box and the predicted bounding box: This represents the distance between the center points of the ground truth bounding box and the predicted bounding box.

[0041] (2) Distance penalty term, as shown below: ; in, It is a distance penalty item; It is an intermediate variable in distance measurement; It is a horizontal distance metric; It is a vertical distance metric; It is the adjustment parameter for distance penalty; It is an overall distance metric; The normalized distance in the horizontal direction is specifically expressed as: , It is the x-coordinate of the center point of the prediction box. It is the width of the smallest bounding rectangle of the ground truth bounding box and the predicted bounding box. It is the x-coordinate of the center of the true bounding box. It is a real frame; It is the normalized distance in the vertical direction, specifically expressed as: , It is the y-coordinate of the center of the true bounding box. It is the ordinate of the center point of the prediction box. It is the height of the smallest bounding rectangle between the ground truth bounding box and the predicted bounding box.

[0042] (3) Shape penalty items, as shown below: ; in, It is a shape penalty item; It is a measure of width difference, specifically expressed as: , It is the width of the prediction box. It is the width of the actual frame; It is a measure of high variability, specifically expressed as: , It is the height of the predicted bounding box. It is the actual height of the bounding box; It is a parameter that adjusts the degree of importance attached to shape loss.

[0043] A lightweight backbone network, an efficient downsampling module, and a detection head achieve accurate detection of tomato flowers and stamens through feature transfer and functional collaboration.

[0044] To verify the detection accuracy and robustness of the DSS-YOLO proposed in this invention, an ablation experiment was set up targeting the core improvements of DSS-YOLO. The experimental training results are shown in Table 1.

[0045] Table 1 Performance Comparison of Ablation Tests

[0046] According to the ablation test results, Experiment 1 used the YOLOv11n benchmark model for detection, and its precision, recall and mAP@50 reached 93.0%, 93.5% and 95.2% respectively. The model weight was 5.2MB, the number of parameters was 2.5M, the computational cost was 6.3GFLOPs, and the inference speed was 57 frames / second.

[0047] In Experiment 2, replacing the backbone network with DWHGNetv2 improved accuracy by 0.8%, reduced model weights, parameters, and computational cost to 3.72MB, 1.8M, and 4.2 GFLOPs, respectively, and increased inference speed to 63 frames per second. This demonstrates that DWHGNetv2 effectively achieves model lightweighting through network structure optimization while simultaneously improving detection accuracy.

[0048] In Experiment 3, after replacing the convolution operation in the detector head with the SCDown module, the accuracy was further improved by 0.6%.

[0049] Experiment 4 showed that the accuracy improved by 0.5% after adopting the SIoU loss function.

[0050] Experiment 5 showcases the overall performance of the DSS-YOLO model of this invention. The DSS-YOLO precision reached 94.2%, a 1.2% improvement over the YOLOv11n baseline model; mAP@0.5 reached 96.2%, a 1% improvement over the YOLOv11n baseline model; and the recall reached 93.5%. Simultaneously, the model weights, number of parameters, and computational cost were reduced by 34.0%, 36.0%, and 35.0%, respectively, while the inference speed increased to 65 frames per second.

[0051] Experimental results show that the DSS-YOLO model effectively alleviates the problem of missed detection caused by factors such as flower stacking, leaf occlusion and large number of targets, and improves the accuracy and robustness of target detection.

[0052] An automated pollination system includes an image acquisition unit for acquiring real-time image data of tomato plants in a greenhouse environment. Data preprocessing unit: used to process real-time image data and perform preliminary feature extraction to obtain model input data; Visual perception unit: integrates the DSS-YOLO tomato flower and stamen detection model, used to identify, classify and locate tomato flowers and stamens in the model input data, and output target category and bounding box coordinates; Pollination execution unit: Based on the bounding box coordinates output by the visual perception unit, it controls the execution structure to complete the precise pollination operation.

[0053] Therefore, the present invention employs the above-mentioned DSS-YOLO tomato flower stamen detection model, which has the following beneficial effects: (1) Through the collaborative design of a lightweight backbone network and an efficient downsampling module, this invention achieves extreme lightweighting while maintaining high accuracy. Experiments show that compared with the benchmark model YOLOv11n, the model weights, number of parameters, and computational cost of this scheme are reduced by 34%, 36%, and 35%, respectively, and the inference speed is increased to 65 frames / second, perfectly meeting the stringent deployment requirements of embedded devices.

[0054] (2) Through the feature fidelity of the efficient downsampling module and the angle-aware mechanism of the SIoU loss function, the present invention significantly enhances the model's ability to detect small-scale, slender targets such as tomato flower stamens. Test results on public datasets show that the model improves the average detection accuracy (mAP@0.5) and precision of flowers and stamens by 1% and 1.2% respectively, effectively reducing false negatives and false positives.

[0055] (3) The multi-scale feature fusion and efficient feature extraction mechanism of this invention enables it to better cope with challenges such as shading, stacking and light changes in the greenhouse environment, and shows more stable detection performance and stronger generalization ability than existing models.

[0056] This invention employs the aforementioned automated pollination system, directly addressing the automated pollination needs in smart agriculture. The system is plug-and-play and adaptable to various scenarios, and has positive practical significance for reducing agricultural labor costs, improving production efficiency, and promoting the intelligent transformation of agriculture.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A DSS-YOLO model for detecting the stamens and pistils of tomato flowers, characterized in that... include: A lightweight backbone network is used to receive raw image data of greenhouse tomato plants, and to extract multi-scale fine-grained features from the images through a low-complexity convolutional structure, outputting an initial feature map. The high-efficiency downsampling module is embedded in the neck network of the DSS-YOLO tomato flower stamen detection model. It is used to reduce the dimensionality and optimize the features of the initial feature map output by the lightweight backbone network, and output the optimized feature map with high discriminative power. The detection head is used to process high-discrimination feature maps. It uses parallel classification and regression branches to output the category probability and bounding box coordinates of tomato flower stamens, respectively, to complete the target detection task.

2. The DSS-YOLO tomato flower stamen detection model according to claim 1, characterized in that, The lightweight backbone network is based on HGNetv2 architecture, and is sequentially connected to the initial feature extraction module HGStem module, five stacked depthwise separable weighted gated (DWHGBlock) modules, spatial pyramid pooling (SPPF) module, channel spatial attention (C2PSA) module, and depthwise separable convolutional downsampling module.

3. The DSS-YOLO tomato flower stamen detection model according to claim 2, characterized in that, The DWHGBlock module consists of five 3×3 depthwise separable convolutions, one 1×1 pointwise convolution, and an ESE attention mechanism layer connected in sequence.

4. The DSS-YOLO tomato flower stamen detection model according to claim 3, characterized in that, Depthwise separable convolution includes channel-wise convolution and pointwise convolution. Channel-wise convolution is used for spatial feature extraction, while pointwise convolution is used for channel dimension adjustment. The number of parameters in depthwise separable convolution is as follows: k×k×c+c×n; Where k is the kernel size; c is the number of input channels; and n is the number of output channels.

5. The DSS-YOLO tomato flower stamen detection model according to claim 1, characterized in that, The high-efficiency downsampling module, as detailed below: The input layer is used to receive the feature maps to be processed. Conv1 consists of Conv2d convolutional layers, BatchNorm2d batch normalization layers, and SiLU activation function, which is used to adjust the number of channels. Conv2 consists of Conv2d convolutional layers and BatchNorm2d batch normalization layers to further extract features. The output layer adds the element-wise values ​​of corresponding positions in the feature maps of Conv1 and Conv2 by element-wise addition to achieve feature reuse; the fused feature map is the final output of the module.

6. The DSS-YOLO tomato flower stamen detection model according to claim 1, characterized in that, The regression branch introduces a bounding box regression loss function, which includes a new angle penalty term. This angle penalty term guides the model to prioritize regression direction by considering the vector angle between the ground truth box and the predicted box. The bounding box regression loss function is shown below: ； in, It is the bounding box regression loss function; It is the intersection-union ratio of the predicted bounding box and the ground truth bounding box; It is a shape penalty term, specifically expressed as: , It is a measure of width difference, specifically expressed as: , It is the width of the prediction box. It is the width of the actual frame; It is a measure of high variability, specifically expressed as: , It is the height of the predicted bounding box. It is the actual height of the bounding box; It is a parameter that adjusts the degree of emphasis placed on shape loss; It is a distance penalty term, specifically represented as: , It is an intermediate variable in the distance metric. It is a horizontal distance metric. It is a vertical distance metric; It is the adjustment parameter for distance penalty. It is an overall distance metric. The normalized distance in the horizontal direction is specifically expressed as: , It is the x-coordinate of the center point of the prediction box. It is the width of the smallest bounding rectangle of the ground truth bounding box and the predicted bounding box. It is the x-coordinate of the center of the true bounding box. It's a real frame. It is the normalized distance in the vertical direction, specifically expressed as: , It is the y-coordinate of the center of the true bounding box. It is the ordinate of the center point of the prediction box. It is the height of the smallest bounding rectangle between the ground truth bounding box and the predicted bounding box.

7. The DSS-YOLO tomato flower stamen detection model according to claim 6, characterized in that, The distance penalty term is obtained based on the newly added angle penalty term. The angle penalty term is as follows: ； in, Angle penalty item; This represents the height difference between the center points of the ground truth bounding box and the predicted bounding box. This represents the distance between the center points of the ground truth bounding box and the predicted bounding box.

8. An automated pollination system, characterized in that, include: Image acquisition unit: used to acquire real-time image data of tomato plants under greenhouse conditions; Data preprocessing unit: used to process real-time image data and perform preliminary feature extraction to obtain model input data; Visual perception unit: integrates the DSS-YOLO tomato flower and stamen detection model as described in claim 1, used to identify, classify and locate tomato flowers and stamens in the model input data, and output target category and bounding box coordinates; Pollination execution unit: Based on the bounding box coordinates output by the visual perception unit, it controls the execution structure to complete the precise pollination operation.

Images