GSEYOLOX-s: An Improved Lightweight Network for Identifying the Severity of Wheat Fusarium Head Blight

Abstract

Fusarium head blight (FHB) is one of the most detrimental wheat diseases. The accurate identification of FHB severity is significant to the sustainable management of FHB and the guarantee of food production and security. A total of 2752 images with five infection levels were collected to establish an FHB severity grading dataset (FHBSGD), and a novel lightweight GSEYOLOX-s was proposed to automatically recognize the severity of FHB. The simple, parameter-free attention module (SimAM) was fused into the CSPDarknet feature extraction network to obtain more representative disease features while avoiding additional parameters. Meanwhile, the ghost convolution of the model head (G-head) was designed to achieve lightweight and speed improvements. Furthermore, the efficient intersection over union (EIoU) loss was employed to accelerate the convergence speed and improve positioning precision. The results indicate that the GSEYOLOX-s model with only 8.06 MB parameters achieved a mean average precision (mAP) of 99.23% and a detection speed of 47 frames per second (FPS), which is the best performance compared with other lightweight models, such as EfficientDet, Mobilenet-YOLOV4, YOLOV7, YOLOX series. The proposed GSEYOLOX-s was deployed on mobile terminals to assist farmers in the real-time identification of the severity of FHB and facilitate the precise management of crop diseases.