Corn Disease Detection Based on an Improved YOLOX-Tiny Network Model

Abstract

In order to detect corn diseases accurately and quickly and reduce the impact of corn diseases on yield and quality, this paper proposes an improved object detection network named YOLOX-Tiny, which fuses convolutional attention module (CBAM), mixup data enhancement strategy, and center IOU loss function. The detection network uses the CSPNet network model as the backbone network and adds the CBAM to the feature pyramid network (FPN) of the structure, which re-assigns the feature maps' weight of different channels to enhance the extraction of deep information from the structure. The performance evaluation and comparison results of the methods show that the improved YOLOX-Tiny object detection network can effectively detect three common corn diseases, such as cercospora grayspot, northern blight, and commonrust. Compared with the traditional neural network models (90.89% of VGG-16, 97.32% of YOLOv4-tiny, 97.85% of YOLOX-Tiny, 97.91% of ResNet-50, and 97.31% of Faster RCNN), the presented improved YOLOX-Tiny network has higher accuracy.