Research on apple leaf pathological detection system based on improved YOLOv7

Apple is one of China's crucial agricultural products, but its yield is often hindered by pests and diseases. To address this issue, a combination of target detection technology and crop leaf pest and disease detection is proposed. This approach aims to practically recognize and analyze the common apple leaf diseases in China. Initially, the YOLOv7 model is selected as the benchmark, and improvements are made to enhance its effectiveness in detecting apple leaf diseases, resulting in the YOLOv7-Me model. The enhancements made to the YOLOv7-Me model include modifications to both the backbone network and the head detection network. Firstly, the Conv2Former is integrated into the backbone network to improve the ELAN structure. This adjustment enhances the target localization accuracy and reduces misdetection. Secondly, the Dyhead target detection head is introduced, which is based on the attention mechanism and combined with CNN convolutional building blocks. This addition enhances the detection of small target diseases. Lastly, the loss function is replaced with Wise-SIoU, which calculates the difference between the prediction results of the semantic segmentation model and the real labels. This change improves the accuracy of detecting apple leaf disease targets. Experimental results demonstrate that the average accuracy of YOLOv7-Me detection is enhanced by 5.01% compared to the baseline model, reaching an accuracy of 85.36%. This improvement signifies better detection outcomes, showcasing the potential of this approach in effectively identifying and managing apple leaf diseases in China.