|
Abstract
|
The adverse effects of hazardous waste on the natural environment and human health are evident. Due to the high quantity of everyday hazardous waste,mutual occlusion, and the complexity of image backgrounds, existing object detection algorithms face challenges in achieving high detection accuracy when detectinghazardous waste. In order to enhance the efficiency of hazardous waste sorting, this paper proposes an improved detection model, DBS-YOLO, based on theYOLOv8n network. This model strikes a balance between accuracy and lightweight design. Firstly, we replace part of the convolution modules in the C2f module withdeformable convolutional networks version 3 (DCNv3) modules, proposing the DC2f module. We apply this module to YOLOv8n, achieving lightweight networkdesign while enhancing the model's ability to extract multi-scale features. Additionally, we introduce a bi-level routing attention (BRA) module to construct a globalreceptive field, capturing long-range dependencies, and improving the model's focus on targets. Finally, we employ a soft non-maximum suppression (Soft-NMS) algorithm to reduce instancesof missed and false detections caused by mutual occlusion, further enhancing model accuracy. Experimental results demonstrate that compared to the originalYOLOv8n network, DBS-YOLO achieves improvements of 2.0% in precision, 1.0% in recall, 2.1% in mAP@0.5, and 4.9% in mAP@0.5:0.95, reaching a mAP@0.5 of95.4%. The model also mitigates instances of missed and false detections. Comparative analyses with SDD, Faster R-CNN, and YOLO series networks confirm theeffectiveness of the proposed model
|