Abstract: To accurately and promptly detect litchi anthracnose among complex natural environmental conditions, a litchi disease recognition method with improved YOLOv7 was proposed．SPPCSPC structure was reconstructed, convolutional layers were pruned, and pooling structure was modified to reduce module complexity and accelerate network convergence speed．To allocate resources reasonably, GAM attention mechanism was introduced．To improve detection accuracy, WIoU loss function was employed．Experimental results indicated that improved YOLOv7 took 0.18 s to detect a single image, with a memory usage of 41.45 MB, and an average accuracy mean of 80.27%．Compared to YOLOv7, memory usage was reduced by 34.5%, detection speed was increased by 60%, and model performance outperformed other models such as Faster R-CNN and YOLOv5．This method provided accurate and rapid detection of litchi disease targets in complex natural environments and unstructured backgrounds, and provided a reference for real-time monitoring research of economic fruit tree leaf diseases．