Lightweight model based on improved YOLOv7 tiny for potato leaf diseases detection

Lightweight model based on improved YOLOv7 tiny for potato leaf diseases detection

Indonesia, a country heavily dependent on agriculture, continues to grow potatoes. However, the presence of plant diseases, manifested by the condition of the leaves, is a significant problem that requires attention. Agriculture offers extensive opportunities to explore computer vision applications, including tasks like object detection. In this paper, we present a method that increases the YOLOv7 tiny model's accuracy to assist farmers in identifying diseases in potato leaves. Our study employed multi-scale and MixUp augmentation techniques to process input images when training using the YOLOv7 tiny model. Based on our experiment, the model can be enhanced using multi-scale training instead of fixed-scale training. After implementing our proposed technique, the mAP metric significantly improved over the original model, achieving a range of 0.94325 to 0.96975 for fixed-scale training and a range of 0.9620 to 0.97525 for multi-scale training with the MixUp approach. In addition, we have developed the YOLOv7 tiny model, which aims to enable seamless use of mobile devices in real-time applications. To assess the current state of potato leaves on land in real-time, we convert the results of our extended model into a compact format called TF Lite. Future potato production can be improved by using these findings to help farmers combat leaf diseases.