Research on object detection methods based on deformable convolutional networks

Object detection is a core problem in the field of computer vision and finds extensive applications in areas such as autonomous driving and security surveillance. Traditional YOLO-based object detection algorithms often struggle with complex backgrounds, varying object shapes, and detecting small objects. To address these issues, this paper proposes an improved YOLO model based on deformable convolutions. Deformable convolutional networks enhance the model’s ability to perceive complex object shapes by introducing spatial deformation capabilities. Experiments were conducted using the publicly available COCO128 dataset and the experimental results show that after introducing deformable convolutions in deeper layers of the model, overall detection accuracy improves, with the mAP50-95 reaching 62.5% when replacing the seventh layer, an increase of 1.2 percentage points compared to the original model. The results indicate that the YOLO model based on deformable convolutions offers certain advantages in handling complex object detection tasks.