A YOLOv5-based Model Enhanced by Dynamic Convolution for Accurate Object Detection

The object detection algorithm based on YOLO has been widely used in many fields, but there are still some shortcomings. As for the classic YOLOv5, it is easy to miss or misdetect when dealing with small objects, resulting in inaccurate detection results. In order to improve the accuracy of object detection, this paper proposes an improved object detection model YOLOv5-DDN. This model replaces the original convolution module in YOLOv5 with dynamic convolution, which enhances the adaptability and representation ability of the model to different input features. In addition, YOLOv5-DDN adopts the DIoU-based non-maximum suppression (DIoU-NMS) technology, which not only considers the overlapping area of the bounding box, but also introduces the distance between the center points of the bounding box as an additional penalty term, so as to more accurately evaluate the similarity between the bounding boxes and avoid the problems of false suppression and repeated detection. We verify the YOLOv5-DDN model on the public dataset GC10-DET. Experimental results show that YOLOv5-DDN outperforms the current state-of-the-art models in terms of detection accuracy.