A multi-scale cucumber disease detection method in natural scenes based on YOLOv5

Plant diseases are the main factors affecting the agricultural production. At present, improving the efficiency of plant disease identification in natural scenarios is a crucial issue. Due to this significance, this study aims at providing an efficient detection method, which is applicable to disease detection in natural scenes. The proposed MTC-YOLOv5n method is based on the YOLOv5 model, which integrates the Coordinate Attention (CA) and Transformer in order to reduce invalid information interference in the background, and combines a Multi-scale training strategy (MS) and feature fusion network to improve the small object detection accuracy. MTC-YOLOv5n is trained and validated on a self-built cucumber disease dataset. The model size and FLOPs are respectively 4.7 MB and 6.1 G, achieving 84.9 % mAP and FPS up to 143. Compared with the advanced single-stage detection model, the experimental results show that MTC-YOLOv5n has higher detection accuracy and speed, smaller computation and model size. In addition, the proposed model is tested under the interference of strong noise conditions such as dense fog, drizzle and dark light, which shows that the model has strong robustness. Finally, the comprehensive experimental results demonstrate that MTC-YOLOv5n is lightweight, efficient and suitable for deployment to mobile terminals for disease detection in natural scenarios.

Automated summary

This study proposes an efficient detection method, MTC-YOLOv5n, based on the YOLOv5 model, for plant disease detection in natural scenarios. The method integrates Coordinate Attention and Transformer, and combines Multi-scale training and feature fusion network to improve small object detection accuracy. Experimental results show that MTC-YOLOv5n has higher detection accuracy and speed, smaller computation and model size, and strong robustness.