Rice leaf tip small target detection and counting method based on global information enhancement

Abstract

The invention discloses a rice leaf tip small target detection and counting method based on global information enhancement, and the method comprises the steps: firstly building a field rice leaf tip image data set as a basis, and constructing an improved YOLO-DCL model; the core innovation of the method lies in that a global context information enhancement module, an inverted residual-cascade grouping attention module and a re-parameterization shared convolution detection head are introduced, key feature expression is enhanced and detection efficiency is optimized through effective fusion of global information, the recognition precision and processing speed of the model on a tiny rice leaf tip target are significantly improved, and the method is suitable for large-scale popularization and application. After training parameters are set, an optimization model is trained, then system performance verification and module optimization are carried out on the model, and a version with the optimal rice leaf tip detection performance is screened out. The technical bottlenecks that small target leaf tips are prone to missing detection and false detection and insufficient in real-time performance under the complex farmland background are effectively overcome, and reliable and efficient technical support is provided for rice growth early-stage yield prediction.

Description

Technical Field

[0001] This paper belongs to the field of computer vision and smart agriculture technology, specifically involving a method for detecting and counting small targets at rice leaf tips based on global information enhancement. It is suitable for the accurate detection and counting of leaf tips during the tillering stage of rice, providing data support for yield prediction.

[0002] Jiangsu Provincial Agricultural Science and Technology Independent Innovation Fund Project (Project No.: CX(22)3016). Background Technology

[0003] As an important food crop, rice yield prediction is directly related to food security and agricultural economy. Most existing rice yield prediction studies focus on the rice maturity stage, with few studies on yield prediction in the early growth stage of rice.

[0004] Patent CN112069985B proposes a high-resolution field image rice panicle detection and counting method based on deep learning, and patent CN112215714B proposes a UAV-based rice panicle detection method and device. These methods input aerial images collected by the UAV into a pre-set MaskR-CNN network model to extract areas with rice panicles. A spatial density clustering algorithm is then used to deduplicate the rice panicle areas in all aerial images before counting the panicles, thus providing support for rice yield prediction. However, both methods rely on mature-stage data with complete panicle morphology. In the early stages of rice growth, panicles have not yet formed or their morphology is not obvious, making panicle detection at this time inaccurate to reflect the rice's growth status. The number of leaf tips during the tillering stage is a key indicator for assessing the number of effective panicles. Traditional manual counting methods rely on experience and suffer from large subjective errors and destructive sampling in complex field environments. In recent years, automatic detection technology based on computer vision has become a research hotspot, but rice leaf tip detection faces the following technical challenges:

[0005] Background interference and target confusion: Rice leaf tips and leaves are highly similar in color and texture, which further increases the difficulty of detection. Patent CN106468543B uses a fixed threshold segmentation algorithm to segment leaves. This method has poor robustness and is prone to misjudging leaf tips as background in overlapping leaf areas. Although the detection method based on improved YOLOv5s used in patent CN117152735A optimizes feature extraction by introducing a hybrid attention mechanism, it has problems such as large computational load, long time consumption, and incorrect assignment of high attention weights to similar color or texture areas in complex backgrounds.

[0006] The target is small and the features are weak: the tip of a rice leaf occupies only 2-3 pixels in the horizontal direction of the image. Its slender shape results in very little effective feature information. Traditional convolutional neural networks are limited by local receptive fields and have difficulty capturing global context information, resulting in a high false negative rate. Although the multi-scale feature fusion method proposed in patent CN111967538A improves the ability to perceive local details through cross-level feature integration, it is insufficient in modeling global spatial relationships and has difficulty distinguishing leaf tips from noise. Summary of the Invention

[0007] To address the aforementioned technical issues, this paper presents a method for detecting and counting small targets at the tips of rice leaves based on global information enhancement. The detailed technical solution is as follows.

[0008] A method for detecting and counting small targets at the tips of rice leaves based on global information augmentation includes the following steps:

[0009] S1: Construct a field image dataset containing rice leaf tip targets, label the leaf tip bounding boxes, and divide it into training set, validation set, and test set according to the proportions;

[0010] S2: Improve the YOLOv8s model structure and construct the YOLO-DCL rice leaf tip small target detection model based on global information enhancement;

[0011] S3: Set training parameters, train the improved model, record the loss curve and evaluation metrics, analyze the training results, and save the optimal model weights.

[0012] S4: Conduct ablation experiments and select the model with the best performance for detecting rice leaf tips;

[0013] S5: The high-resolution rice paddy image to be detected is cropped into several sub-images of appropriate size. These sub-images are input into the trained model. The model outputs the number of leaf tip prediction boxes. After counting and deduplicating the leaf tip prediction boxes, the result of multiplying the number of prediction boxes by the growth stage counting compensation factor is the number of rice leaf tips.

[0014] In step S1, a camera is used to collect data from the rice paddy. The camera's height and tilt angle are adjusted to optimize the shooting effect. This adjustment helps to reduce the adverse effects caused by reflected light and shadows, ensuring the acquisition of high-quality rice field images. The rice leaf tips are labeled using the Labelme annotation tool. The labeled data is divided into training set, validation set, and test set. Data augmentation techniques are used to finally generate a high-quality rice leaf tip image dataset.

[0015] In step S2, the YOLO-DCL model based on global information enhancement includes: a backbone network that integrates inverted residual-cascaded grouping attention modules, a multi-scale feature pyramid network with added global context information enhancement modules, and a detection head that generates prediction boxes based on reparameterization and shared convolution.

[0016] The backbone network includes: the original YOLOv8 Conv module, C2f module, SPPF module, and the proposed inverted residual-cascaded grouping attention module;

[0017] The inverted residual-cascaded grouping attention module proposed in the backbone network includes: generating multi-scale features at different levels by channel expansion, then dividing the features into multiple groups, using a cascaded structure of multi-stage grouping attention heads, independently generating queries, keys, and values ​​for each group and calculating attention weights, then enhancing the spatial context of the keys through deep convolution, and superimposing the output of the preceding head into the input of the subsequent head to gradually integrate multi-scale information, then concatenating the outputs of all heads to form a progressive feature refinement process, and finally the channel compression operation to select the channel components most sensitive to target localization and semantic discrimination, forming a closed-loop feature processing process of expansion-refinement-compression;

[0018] The multi-scale feature pyramid network includes: the original YOLOv8 Conv module, Concat module, Upsample module, and the proposed global context information enhancement module;

[0019] The global context information enhancement module proposed in the multi-scale feature pyramid network consists of a double convolutional layer, a dual-branch feature block mixing module based on depthwise separable convolution, and a channel interaction module.

[0020] The dual-branch feature block fusion module based on depthwise separable convolution processes each channel of the input feature map independently using the separated dual-branch depthwise convolution, establishing long-distance dependencies between regions within each channel. Then, a dynamic feature fusion module adaptively integrates the features to generate a feature representation that is sensitive to both local details and global structure. More specifically, the dynamic feature fusion module uses dual-channel 1×1 convolution to generate a weight map and combines it with a normalized exponential function to dynamically allocate weights, allowing global and local features to be weighted differently based on semantic importance. Finally, the weighted result is superimposed on the original input features to form a feature representation output that is sensitive to both local details and global structure. The specific process can be represented as follows:

[0021]

[0022] F2 = Conv 1×1 ·ReLU(F1)+b2,F2∈R B×2×H×W

[0023] w G ,w L =Softmax(F2,dim=1)

[0024] Output = w G Q+w L V+X

[0025] In the formula, X represents the input feature map, and Conv 1×1 This represents a convolution with a 1×1 kernel, b1 and b2 represent the convolution bias, B, C, H, and W represent the dimensions of the output features, B represents the batch size, C represents the number of input channels, H and W represent the height and width of the feature map, ReLU represents the ReLU non-linear activation function, and w G w represents the dynamic global feature weights. L represents the dynamic local feature weights, Softmax represents the Softmax activation function, dim=1 indicates that the Softmax function is applied to channel C, and Q and V represent the local context encoding and channel importance weights generated by the network, respectively.

[0026] The channel interaction module uses two layers of pointwise convolution and the GELU activation function between the two convolution layers to combine features of different channels, and SE attention to capture channel-level global statistical information, strengthen important semantic channels, and finally realize information interaction between different channels.

[0027] The prediction box generation detection head based on reparameterization and shared convolution includes: forming a deep feature compensation mechanism by stacking two shared convolutional layers to effectively alleviate the loss of shallow feature information caused by parameter sharing; and using reparameterization technology to merge multi-branch topologies containing 3×3 convolutions, 1×1 convolutions, and identity mappings, as shown in the calculation formula below. Group normalization and learnable scale function layers are introduced to optimize feature distribution.

[0028]

[0029]

[0030] In the formula α i W represents the normalization scaling factor. i b represents the kernel weight of the i-th branch. i μ represents the kernel bias of the i-th branch. i β represents the normalized mean. i This represents the normalized offset.

[0031] Step S5 includes considering that the impact of leaf morphology changes on counting accuracy during different growth cycles of rice is dynamic. A growth stage counting compensation factor is added according to the rice growth time. The counting strategy is adaptively adjusted in combination with growth time information to compensate for the detection deviation caused by leaf morphology differences, so that the detection results are more consistent with the actual number of leaf tips.

[0032] Beneficial effects:

[0033] This invention addresses the technical challenges of rice leaf tip detection by specifically improving the original YOLOv8s model. A cascaded grouped attention module effectively captures subtle leaf tip features through progressive feature refinement. Combined with global information enhancement and dynamic feature fusion, the model's ability to distinguish similar textured regions is improved, enabling precise leaf tip localization and reduced background interference in complex field environments. The lightweight design, through reparameterization and shared convolution techniques, significantly reduces model complexity while maintaining high accuracy, achieving efficient inference on mobile devices. This provides technical support for early-stage rice yield prediction and is of significant value in promoting the development of smart agriculture. Attached Figure Description

[0034] Figure 1 A flowchart of a method for detecting and counting small targets at the tips of rice leaves based on global information enhancement;

[0035] Figure 2 This is a dataset image display for a method of detecting and counting small targets at the tips of rice leaves based on global information enhancement;

[0036] Figure 3 The YOLO-DCL model architecture is used in a method for detecting and counting small targets at the tips of rice leaves based on global information enhancement.

[0037] Figure 4 This is a structural diagram of the inverted residual-cascaded grouping attention module in the YOLOv8s model replaced in a method for detecting and counting small targets at the tips of rice leaves based on global information enhancement.

[0038] Figure 5 This is a structural diagram of the global context information enhancement module that replaces the YOLOv8s model in a method for detecting and counting small targets at the tips of rice leaves based on global information enhancement.

[0039] Figure 6 This is a diagram of the reparameterized shared convolutional detection head structure in a rice leaf tip small target detection and counting method based on global information enhancement, which replaces the YOLOv8s model.

[0040] Figure 7 A flowchart illustrating the training approach used to train the YOLO-DCL model for rice leaf tip detection proposed in this invention.

[0041] Figure 8 This image shows the detection results of rice leaf tips using a method for detecting and counting small targets at rice leaf tips based on global information enhancement. Detailed Implementation

[0042] To enhance understanding of the present invention, the invention will be further described in detail below with reference to embodiments and accompanying drawings. These embodiments are only for explaining the invention and do not constitute a limitation on the scope of protection of the invention.

[0043] like Figure 1 , 2 As shown in points 3, 4, 5, 6, 7, and 8:

[0044] refer to Figure 1 This invention provides a technical solution: a method for detecting small targets at the tips of rice leaves based on global information enhancement, the specific steps of which include:

[0045] S1: Label the leaf tips of rice in the field to construct a dataset of rice leaf tip images in the field, and use the Labelme annotation tool to label the rice leaf tips;

[0046] Specifically, it includes the following sub-steps:

[0047] S1.1: Data was collected from the rice paddies using a camera. The camera's height and tilt angle were adjusted to optimize the shooting effect. This adjustment helps to reduce the adverse effects caused by reflected light and shadows, ensuring the acquisition of high-quality images of the rice fields. The original resolution of the acquired images was 5400×4050 pixels. The acquired images were then cropped into several appropriately sized sub-images with a resolution of 1664×1248. The sub-images are shown below. Figure 2 As shown;

[0048] S1.2: Use the Labelme annotation tool to annotate the leaf tips of the rice field images captured by the camera, generate label files for training the YOLO model, and divide the annotated data into training set, validation set and test set according to the ratio of 9:0.5:0.5;

[0049] S1.3: Image data augmentation techniques are applied to the images to effectively enhance the diversity of the dataset. Specific operations include various transformation methods such as translation, scaling, mirroring, and mosaic processing. These techniques enhance the model's adaptability to different environments by simulating various complex situations that may occur in actual agricultural scenarios, and ultimately generate a high-quality dataset of rice leaf tips in the field.

[0050] S2: Improve the YOLOv8s model network architecture by constructing a YOLO-DCL model based on global information augmentation. The YOLO-DCL model is as follows: Figure 3As shown;

[0051] Specifically, it includes the following sub-steps:

[0052] S2.1: Add a fused inverted residual-cascaded grouped attention module to the backbone network of the YOLOv8s model, with the structure as follows: Figure 4 As shown;

[0053] The inverted residual-cascaded grouping attention module added to the backbone network includes: generating multi-scale features at different levels by channel expansion, then dividing the features into multiple groups, using a cascaded structure of multi-stage grouping attention heads, independently generating queries, keys, and values ​​for each group and calculating attention weights, then enhancing the spatial context of the keys through deep convolution, and superimposing the output of the preceding head into the input of the subsequent head to gradually integrate multi-scale information, then concatenating the outputs of all heads to form a progressive feature refinement process, and finally channel compression to select the channel components most sensitive to target localization and semantic discrimination, forming a closed-loop feature processing process of expansion-refinement-compression;

[0054] S2.2: The proposed global context information enhancement module is used in the multi-scale feature pyramid network of the YOLOv8s model, with the structure as follows: Figure 5 As shown;

[0055] The global context information enhancement module in the multi-scale feature pyramid network consists of a double convolutional layer, a dual-branch feature block mixing module based on depthwise separable convolution, and a channel interaction module.

[0056] The dual-branch feature block mixing module based on depthwise separable convolution processes each channel of the input feature map independently using the separated dual-branch depthwise convolution, establishing long-distance dependencies between regions within each channel. Branch 1 uses a 3×3 depthwise convolution with no padding and a stride of 1, outputting a feature map X3 for extracting global structural features. Branch 2 uses a 1×1 depthwise convolution, which is equivalent to a channel-wise identity transformation, outputting a feature map X1 for preserving local detail features.

[0057] The output feature maps X3 and X1 are then adaptively integrated by a dynamic feature fusion module. In the dynamic fusion module, X3 is equal to the channel importance weight V, and X1 is the local context encoding Q. Finally, a feature representation with both local detail sensitivity and global structure awareness is generated. More specifically:

[0058] The dynamic feature fusion module first uses the formula F1 = Conv 1×1 (X)+b1, Perform channel-reduction convolution, using 1×1 convolution to reduce the number of input feature channels to 1 / 16 of the original number of channels, initialize the bias b1 to 0, and output F1;

[0059] Then, using the formula F2 = Conv 1×1 ·ReLU(F1)+b2,F2∈R B×2×H×W Nonlinear activation and weight generation are performed. After applying the ReLU activation function to F1, a 2-channel weight map is generated through a 1×1 convolution. Here, Conv... 1×1 It contains two 1×1 convolution kernels, with the bias b2 initialized to 0, and outputs F2;

[0060] Then, using formula w G ,w L =Softmax(F2,dim=1) performs Softmax normalization on F2 along the channel dimension (dim=1) to generate the global weight w. G and local weight w L Both have the same dimension and satisfy w pixel by pixel. G +w L =1;

[0061] Finally, the formula Output = w is used. G Q+w L V+X fuses the global feature V and the local feature Q according to weights, and then superimposes the original input X to obtain the output feature map;

[0062] The channel interaction module uses two layers of pointwise convolution and the GELU activation function between the two convolution layers to combine features of different channels, and SE attention to capture channel-level global statistical information, strengthen important semantic channels, and finally realize information interaction between different channels.

[0063] S2.3: In the detection head part of the YOLOv8s model, the original detection head is replaced with a prediction box generation detection head based on reparameterization and shared convolution, with the structure as follows: Figure 6 As shown;

[0064] The prediction box generation detection head based on reparameterization and shared convolution includes: forming a deep feature compensation mechanism by stacking two shared convolutional layers to effectively alleviate the loss of shallow feature information caused by parameter sharing. The kernel size of the first shared convolutional layer is 3×3 with a stride of 1, keeping the spatial size of the feature map unchanged. The dimension of the second shared convolutional layer is the same as that of the first shared convolutional layer.

[0065] Furthermore, a reparameterization technique is employed to merge multi-branch topologies containing 3×3 convolutions, 1×1 convolutions, and identity mappings. Through the linear equivalence transformation property of batch normalization, multi-branch convolution operations are merged into a single 3×3 convolution kernel during the inference stage. Firstly, according to the formula... For each branch, apply its own normalization scaling factor α iAdjust the convolution kernel weights, then perform a weighted sum of the adjusted kernel weights to obtain the merged kernel weight W, and then apply the formula... To merge convolution kernel biases, first calculate the bias of each branch, specifically b in each branch. i -μ i The value of b i μ represents the kernel bias of the i-th branch. i This represents the normalized mean, which is then multiplied by the corresponding normalization scaling factor α. i In addition, the normalized offset β i Finally, the results of the three branches are added together to obtain the merged convolution kernel bias b. In the inference stage, the merged convolution kernel weight W and convolution kernel bias b are equivalent to a multi-branch structure of 3×3 convolution, 1×1 convolution and identity mapping. Then, group normalization and learnable scale function layers are introduced to optimize the feature distribution, and the final feature map used for generating prediction boxes is obtained.

[0066] S3: Set training parameters, train the improved YOLO-DGL model, record the loss curve and evaluation metrics, analyze the training results, and save the optimal model weights.

[0067] Specifically, it includes the following sub-steps:

[0068] S3.1: The model training rounds are set to 100 rounds, the input image size is 1664×1248, the batch size is 4, the SGD optimizer is selected to update the model weights during model training, and the learning rate is set to 0.01. In order to ensure the stability in the later stage of training and avoid large fluctuations when the model is close to the optimal solution, a learning rate cosine annealing decay strategy is adopted, and the learning rate is gradually reduced as the number of training rounds increases.

[0069] S3.2: Import YOLOv8s pre-trained model weights to start model training. Using the pre-trained model to initialize the new model can speed up the convergence speed and improve the generalization ability of the model.

[0070] S3.3: Calculate the validation set mAP@0.5, recall, and precision in each round. Record the loss curve and evaluation metrics through TensorBoard, save the best model weights and ONNX format conversion, and trigger the early stopping mechanism when there is no performance improvement for 10 consecutive epochs.

[0071] S4: Verify the product performance and optimize the modules of the model proposed in this invention, and select the model with the best performance for rice leaf tip detection.

[0072] Specifically, it includes the following sub-steps:

[0073] S4.1: The YOLOv8s-DCL model of this invention proposes three improved modules, all of which can be integrated into the YOLOv8s model independently. In order to verify the performance contribution of each functional module to the rice leaf tip detection method, the optimal configuration scheme is determined by comparing the detection capabilities and resource efficiency of different module combinations.

[0074] S4.2: Table 1 shows the core performance indicators of each version of the system. Precision is the accuracy. The higher the accuracy, the lower the false positive rate of the model. Recall is the recall. The higher the recall, the lower the false negative rate of the model. mAP@0.5 is the average precision when the cross-union threshold for leaf tip detection is 0.5. Params is the number of model parameters. FLOPs is the computational cost of the model. FPS is the number of rice images that the model can process per second.

[0075] Table 1 Comparison of Module Combination Performance

[0076]

[0077] As shown in Table 1, all three improved modules proposed in this paper improve the detection performance of the YOLOv8 algorithm. Replacing the C2f module of the multi-scale feature pyramid network with the global context information enhancement module significantly improves model performance, reducing the number of parameters by 12.5%, the computational cost by 9.8%, and the average precision by 1.8% compared to the YOLOv8s model. Secondly, the introduction of a reparameterized shared convolutional detector head reduces the number of model parameters by 15.1% and the required floating-point operations by 9.1%, while maintaining an average precision roughly equivalent to the YOLOv8s model. The introduction of inverted residual-cascaded grouping further enhances the performance. When the attention module was introduced, the model's average precision improved by 0.6%, recall improved by 2.4%, and the model's lightweighting was also improved. This proves that the inverted residual-cascaded grouping attention module can effectively improve the model's ability to locate small targets, thereby reducing the model's false negative rate, and it also has a certain degree of lightweight performance. When the global context information enhancement module, the inverted residual-cascaded grouping attention module, and the reparameterized shared convolutional detection head are introduced at the same time, the model's overall performance reaches the highest level, the resource consumption is compressed to 7.02M, the computational efficiency is improved by 22.1%, and excellent detection accuracy is maintained. It can achieve high-performance rice leaf tip small target detection effect on resource-constrained mobile devices.

[0078] S5: The high-resolution rice paddy image to be detected is cropped into several sub-images of appropriate size. These sub-images are then input into the trained model. The model outputs the number of leaf tip prediction boxes. The detection results are as follows: Figure 8As shown, after counting and deduplicating the leaf tip prediction boxes, the result of multiplying the number of prediction boxes by the growth stage count compensation factor is the number of rice leaf tips. The growth stage count compensation factor is shown in Table 2 below:

[0079] Table 2 Compensation Factors for Growth Stage Counting

[0080] Growth time Count compensation factor reason 7-10 days after transplanting Count value × 0.8 Tillering not fully developed 11-20 days after transplanting Count value × 1.0 -- 21-25 days after transplanting Count value × 1.2 Considering leaf overlap and shading

[0081] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting and counting small targets at the tips of rice leaves based on global information augmentation, characterized in that, include: S1: Construct a field image dataset containing rice leaf tip targets, label the leaf tip bounding boxes, and divide it into training set, validation set, and test set according to the proportions; S2: Improve the YOLOv8s model structure and construct the YOLO-DCL rice leaf tip small target detection model based on global information enhancement; S3: Set training parameters, train the improved model, record the loss curve and evaluation metrics, analyze the training results, and save the optimal model weights. S4: Verify the product performance and optimize the modules of the model proposed in this invention, and select the model with the best performance for rice leaf tip detection. S5: Input the rice image to be detected into the trained model. The model outputs the number of leaf tip prediction boxes. The result of multiplying the number of prediction boxes by the growth stage counting compensation factor is the number of rice leaf tips.

2. The method for detecting and counting small targets at the tips of rice leaves based on global information enhancement according to claim 1, characterized in that, Step S1 includes: S1.1: Use a camera to collect data from rice fields, adjust the camera's height and tilt angle to optimize the shooting effect. This adjustment helps to reduce the adverse effects caused by reflected light and shadows, ensuring the acquisition of high-quality rice field images. Use the Labelme annotation tool to annotate the tips of rice leaves. S1.2: Divide the labeled data into training set, validation set, and test set, and use the divided dataset to train the target detection model based on global information enhancement.

3. The method for detecting and counting small targets at the tips of rice leaves based on global information enhancement according to claim 1, characterized in that, Step S2 includes: S2.1: Construct a global context information enhancement module, which includes a feature block mixing module and a channel interaction module. The global context information enhancement module generates a weight map through dual-channel 1×1 convolution and achieves dynamic feature fusion by combining a normalized exponential function. The calculation formula is as follows: F2=Conv 1×1 ·ReLU(F1)+b2,F2∈R B×2×H×W w G w L =Softmax(F2,dim=1) Output=w G Q+w L V+X In the formula, X represents the input feature map, and Conv 1×1 This represents a convolution with a 1×1 kernel. b1 and b2 represent convolution biases, B, C, H, and W represent the dimensions of the output features, B represents the batch size, C represents the number of input channels, H and W represent the height and width of the feature map, ReLU represents the ReLU non-linear activation function, and w G w represents the dynamic global feature weights. L represents the dynamic local feature weights, Softmax represents the Softmax activation function, dim=1 indicates that the Softmax function is applied to channel C, and Q and V represent the local context encoding and channel importance weights generated by the network, respectively. S2.2: Construct an inverted residual-cascaded group attention module. Combining the inverted residual block and the cascaded attention mechanism, a hierarchical feature refinement closed loop is formed. First, channel expansion is used to generate multi-scale features at different levels. The cascaded group attention integrates these multi-scale features through cross-head information transmission, so that each attention head can inherit the spatial context information of the previous calculation, realizing a progressive feature refinement process. Finally, the channel compression operation selects the channel components that are most sensitive to target localization and semantic discrimination, forming a closed-loop feature processing process of expansion-refinement-compression. S2.3: Construct a reparameterized shared convolutional detector head, merging multi-branch convolutions through reparameterization techniques. The calculation formula is shown below. Group normalization and learnable scale function layers are introduced to optimize feature distribution. In the formula α i W represents the normalization scaling factor. i b represents the convolution kernel weight of the i-th branch. i μ represents the kernel bias of the i-th branch. i β represents the normalized mean. i This represents the normalized offset.

4. The method for detecting and counting small targets at the tips of rice leaves based on global information enhancement according to claim 1, characterized in that, Step S5 includes: considering that the impact of leaf morphology changes on counting accuracy during different growth cycles of rice is dynamic, a growth stage counting compensation factor is added according to the rice growth time.

5. A method for detecting and counting small targets at the tips of rice leaves based on global information enhancement, as described in any one of claims 1, 2, 3, and 4, characterized in that, The detection method for the aforementioned model includes: A rice leaf tip dataset was constructed, containing high-resolution images and labeled samples from multiple scenarios. The model is used for feature extraction, fusion, and detection to achieve accurate localization and counting of leaf tip targets.

Images