改进YOLOv5su模型检测桃树缩叶病

为实现自然环境下桃树缩叶病的检测，该研究提出了一种基于YOLOv5su的桃树缩叶病识别改进模型DLL-YOLOv5su。首先，针对桃树缩叶病目标特征变化较大的问题，在骨干网络最后一层C3模块中加入可变形自注意力模块（deformable attention，DA），使模型更加关注目标区域，降低背景对模型的影响，提高模型在复杂背景下的拟合能力。其次在SPPF（fast spatial pyramid pooling）模块中引入LSKA（large separable kernel attention）结构，大核卷积增大了模型的感受野，使模型能够关注更多信息。最后，提出了LAWD（lightweight adaptive weighted downsampling）模块，使用轻量化的下采样结构替换卷积模块，减少计算开销。在桃树缩叶病数据集上进行试验，结果显示，DLL-YOLOv5su模型权重大小为17.6MB，检测速度为83帧/s。识别准确率P、召回率R和平均精度均值mAP₅₀分别达到了80.7%、73.1%和80.4%，相较于原始YOLOv5su分别提高了4.2、2.4和4.3个百分点。与YOLOv3-tiny、Faster R-CNN、YOLOv7和YOLOv8相比mAP₅₀分别高出了28.5、11.8、2.1和4.1个百分点。改进模型识别精度高，误检、漏检率低，检测速度满足实时检测的要求，可以为桃树缩叶病的实时监测和预警提供参考。

Improved YOLOv5su model for detecting peach leaf curl disease

Peach trees are susceptible to leaf curl disease in humid regions in spring, leading to a reduction in peach production. Current monitoring cannot timely and accurately identify the leaf curl disease in large-scale orchards. Therefore, intelligent detection can be expected for the peach tree diseases using deep learning, in order to minimize the pesticide application. However, most research has relied on individual-picked leaves in disease identification. It is still challenging to apply to disease detection in real-world scenarios. The primary research can also emphasize to differentiation among various disease types. Yet it is lacking in the accurate occurrence of leaf disease. It is also difficult to identify the peach leaf curl, due to the complex background of disease images and the small size of the early target. In this study, a DLL-YOLOv5su model was introduced to detect peach leaf curl disease in natural environments. Firstly, a deformable attention module was integrated into the backbone network of YOLOv5su. The C3-DA module was formed to recognize the targets with significant size variations, in order to improve the feature fitting of the model. Additionally, a Large Separable Kernel Attention (LSKA) module was incorporated into the Spatial Pyramid Pooling Fusion (SPPF) layer to expand the receptive field of the network. The sensitivity to large targets was enhanced without complexity. Lastly, a lightweight adaptive weighted downsampling (LAWD) module was proposed on the receptive field attention convolution (RFAConv), in order to replace the convolutional modules within the network for feature extraction. Both lightweight and accuracy were achieved concurrently. The dataset was also built for the peach leaf curl disease, including the early, middle, and late stages. 1520 images were then collected from the peach orchard of Chongqing Academy of Agricultural Sciences, China. Among them, the number of pictures in the training set increased from 1105 to 6630 after data augmentation. Subsequently, ablation experiments were conducted to validate the improvement under the same device environment. Each improvement was observed. The recall rate increased by 3.5 percentage points under the C3-DA module, compared with the initial model, while the mean average precision (mAP₅₀) increased by 2.3 percentage points; The accuracy and average precision (mAP₅₀) increased by 2.5 and 2.3 percentage points, respectively, after the LSKA module was introduced; The model size decreased by 18.9 percentage points, after the LAWD was used to replace some CBS modules. The ablation experiments revealed that the improved DLL-YOLOv5su model was achieved with a detection accuracy of 80.7%, a recall rate of 73.1%, and an average precision mean average precision (mAP₅₀) of 80.4% on the test set of peach leaf curl disease. Compared with the initial model, these metrics increased by 4.2, 2.4, and 4.3 percentage points, respectively. The frame rate (FPS) of 83 frames per second fully met the requirements of real-time detection for peach leaf curl. A comparison experiment was designed to compare the detection performance of DLL-YOLOV5su with the current mainstream model of target detection. The performance of the DLL-YOLOv5su model outperformed the most mainstream models, including YOLOv3-tiny, YOLOv7, Faster R-CNN, and YOLOv8, with higher accuracy and precision. The recall rate was only 1.5 percentage points lower than YOLOv7, but the model size was only 23.5% of YOLOv7''s. In summary, the improved DLL-YOLOv5su model was achieved in the real-time and accurate detection of peach leaf curl occurrences, thereby enhancing the efficiency of smart peach orchards. Resource allocation was also optimized to facilitate the precise application of pesticides, particularly for the high crop yields and healthy peach trees.