DB-YOLOv8 Model: Object Detection in Complex Scenarios

The improved DB-YOLOv8 model aims to address the issues of missed detection and false detection that are prone to occur in object detection tasks under complex scenarios, particu-larly in the case of overlapping objects and multiple identical ob-jects. It targets the limitations of the YOLO backbone network caused by fixed kernels that lead to restricted receptive fields and insufficient multi-scale feature learning capabilities due to the in-ability of the spatial and channel attention fusion mechanism to adapt to the input data feature distribution. The DB- YOLOv8 model introduces the deformable convolution C2f (D _ C2f) module in the backbone network to achieve adaptive adjustment of the network's receptive field. Additionally, the BiFPN structure is in-troduced in the Neck network, allowing information to propagate bi-directionally between different resolution levels, thereby better fusing multi-scale information. Experimental results demonstrate the outstanding performance of the DB-YOLOv8 model on the VOC and COCO datasets. Validation confirms that its mAP50 and mAP50-95 metrics are improved by 3.3% and 4.7% compared to the YOLOv8 model, respectively.