Using an Improved YOLOv5 Network for the Automatic Detection of Silicon on Wheat Straw Epidermis of Micrographs

Abstract

The silicon on wheat straw epidermis is an obstacle to its resource utilization, and pretreated methods should be applied to remove this structure. Due to the difficulty in detecting the silicon on wheat straw epidermis, judging the efficiency of pretreatment is still a challenging task. In this study, an automatic detection method based on YOLOv5s was proposed to detect the silicon on wheat straw epidermis of micrographs. To improve the efficiency of the network, the input was modified, the inverted residual module, the pointwise convolution and the attention mechanism were added, the Focus module was cut off. 3350 micrographs of wheat straw epidermis were collected for training and testing. The training results showed that the improved model can efficiently detect silicon on wheat straw epidermis of micrographs, and had the highest mAP (mean Average Precision) of 99.72% among five state-of-the-art comparison models including RetinaNet, SSD, YOLOv4tiny, YOLOv4 and YOLOv5s. The weight of the improved model was 11.7 M, indicating that it can be transplanted to mobile devices. The improved model showed good robustness under different imaging conditions. All the results indicated that the improved model could detect the silicon on wheat straw epidermis of micrographs accurately and efficiently.