EFDet: An efficient detection method for cucumber disease under natural complex environments

Abstract Improving the application capability of the disease detection model is a key issue in the field of agricultural informatization. The complex backgrounds, image diversity, and model complexity are the main factors that affect the realization of automatic disease recognition. This study constructs an efficient detection model (EFDet), which mainly consists of the efficient backbone network, a feature fusion module, and a predictor. EFDet improves the detection effect for cucumber leaves in complex backgrounds by fusing feature maps at different levels. We collected three category cucumber leaves including downy mildew, bacterial angular spot, and health to construct the cucumber disease dataset. It contains 7,488 images with three complexity levels for model training and evaluation. YOLO V3-V5, EfficientDet-D1, YOLO V3-ASFF, and other six detection models as the comparison models, we verify the EFDet performance in terms of model size , FLOPs , and mAP. Experimental results show that EFDet has strong robustness for cucumber disease leaf in complex environments. It also has smaller parameters and calculations that are suitable for actual applications.