Real-time object detection method of melon leaf diseases under complex background in greenhouse

Early disease detection in greenhouses is an important part of integrated disease management in modern agriculture. A real-time object detection method of melon leaf disease, Pruned-YOLO v5s+Shuffle (PYSS) is proposed in this research. First, for enhancing the feature extraction capability, the backbone of the YOLO v5s is reconstructed with ShuffleNet v2 Inverted Residual block. Then, to further downsize the model, the channel pruning method is used to prune and fine-tune the sparsely trained model. Finally, Pruned-YOLO v5s+Shuffle model is deployed to Jetson Nano, and the real-time performance is confirmed in melon greenhouses. The experimental results show that the proposed model has 93.2% and 98.2% mAP@0.5 for melon ( Cucumis melon. L) powdery mildew and melon real leaves, respectively. Compared with YOLO v5s, the performance of our proposed model is improved 6.2% and 6.4% in the term of mAP@0.5 and precision, respectively. The model size and inference time are reduced 85% and 7.5%. In addition, the PYSS demonstrates the higher detection precision and faster inference speed in the comparison of YOLO v3, Faster R-CNN, RetinaNet, Cascade R-CNN, YOLO v4 and YOLO v5s. Being deployed to Jetson Nano, the detection results are displayed on the monitor in real time: mAP@0.5 is 96.7%, the model size is 1.1 MB, and the inference time is 13.8 ms.