A Multi-objective Apple Leaf Disease Detection Algorithm Based on Improved TPH-YOLOV5

To address the challenge of apple leaf disease target recognition in complex environments, we proposed an improved apple leaf disease detection algorithm based on TPH-YOLOV5. We filtered, supplemented, and enhanced the apple leaf disease dataset. By introducing the SimAM attention module and adopting the MobileNetV3 module to reduce network parameters, we improved the TPH-YOLOV5 network. To further reduce parameters, we removed the prediction head for tiny objects from TPH-YOLOV5, resulting in the specialized TPH-YOLOV5-MobileNetV3 network with only 4,537,842 parameters. After 500 iterations on the training set, the improved model achieved precision, recall, F1 score, and mean average precision (mAP) of 91.41%, 91.94%, 91.67%, and 94.26%, respectively. Additionally, the trained model of this network had a size of 11.8 MB, with a testing speed of 17.05 ms per image. Comparing with YOLOV5, YOLOV5-MobileNetV3, and TPH-YOLOV5, our experiments demonstrated that the improved apple leaf disease detection network significantly reduced model parameters and detection time while maintaining recognition accuracy. This network balances recognition effectiveness and speed, making it suitable for practical applications and providing theoretical support for the rapid, non-destructive identification of apple leaf diseases.