YOLOv5s-CBAM-DMLHead: A lightweight identification algorithm for weedy rice (Oryza sativa f. spontanea) based on improved YOLOv5

Abstract

Rice (Oryza sativa L.) is one of the essential food sources for people, with rice farms producing about 480 million tons of refined rice annually. In recent years, with the development of strongly dominant hybrid rice technology, weedy rice (Oryza sativa f. spontanea) poses a serious threat to rice production and reduces the yield and quality of rice. At the seed production stage, weedy rice significantly reduces the purity of harvested seeds and even causes seed production failure. To ensure rice production and solve the challenge of rice seed production purity. In this study, we constructed WeedyRice5, a conventional weedy rice image dataset collected by a camera in a natural context, containing 6436 weedy rice images of five types of rice at the booting stage in the study area. We construct the target detection model for WeedyRice5 using the advanced YOLOv5s method, where the model mAP@0.5 was 98.2%. To improve model performance, a new YOLOv5s-CBAM-DMLHead method based on YOLOv5s is proposed to reduce the model computation and detection time. First, the detection and recognition rates are improved by replacing all C3 modules in the backbone network with CBAM attention mechanism modules. Then, by analyzing the target size in the dataset, we discover that there are more small targets and fewer large targets. Therefore, we removed the large and medium target detection feature maps from the neck network to further reduce the computational cost of the model. Finally, the online training set was expanded using the mosaic online data enhancement method, thus improving the generalizability and robustness of the model, and preventing overfitting. The experimental results demonstrate the effectiveness of the proposed model. The results show that the mAP@0.5 is 98.9%, the average detection time is 4 ms, and the computational demand is reduced by about 28% compared with the original YOLOv5s model. These results meet the requirements of precision, real-time, and lightweight features. In addition, the method inherits the lightweight features of YOLOv5 and has excellent application prospects in rice variety selection, large-area seed production, and regional trials.