YOLO-WDNet: A lightweight and accurate model for weeds detection in cotton field

Accurate and rapid weed detection is important step for crop precision management. Existing deep learning based weed detection algorithms often have huge number of parameters occupying operation resources, which makes it difficult to apply in the actual environment. In addition, detection of weeds in unstructured environments still remains challenging due to the dimensional change, illumination variation, and occlusion or adhesion phenomenon. Here, we constructed a cotton and weed detection dataset (CWDD) including 4,694 images with 8,969 boxes from real scenario. A lightweight YOLO v5 based cotton & weed detection network (YOLO-WDNet) trained on the CWDD images was developed, aiming to solve existing weed identifying problems. We further refactored the feature extraction module by employing shuffleNet v2, three-level BiFPN and novel parallel hybrid attention mechanism (PHAM). And the newly EIOU loss function was designed to improve distinguishing ability for detecting multi-scale and overlapping targets. The performance of novel YOLO-WDNet was studied comprehensively and compared with the state-of-the-art (SOTA) models. The parameter number of YOLO-WDNet reduced by 82.3%, the model size reduced by 91.6%, and floating point of operations reduced by 82.3%. The detection mAP_0.5 of proposed method improved by 9.1%, and the inference time decreased by 57.14%. YOLOWDNet also outperformed the advanced SOTA algorithms for weeds detection. We deployed the trained YOLOWDNet model to the mobile ground robot to verify the detection performance in complex unstructured cotton farmland. Our study shows that the YOLO-WDNet has clear advantages in terms of accuracy, efficiency and lightweight performance. It also provides a paradigm for the lightweight design and application of visual perception models in agricultural scenes, which can be potentially used for the automatic detection and controlling of field weeds on mobile spraying equipment.