| Object detection is one of the most important research topics in computer vision,with wide-ranging research and practical applications in achieving automation,improving efficiency,and promoting technological development.Nowadays,large and medium-sized target detection has achieved relatively good detection results because the target is relatively easy to detect.However,small targets have problems such as low resolution,dense distribution and easy overlapping,making detection more difficult compared to large object detection.In response to the above challenges,this paper uses YOLOv8 as the benchmark model,and optimizes its IoU calculation methods,neck andbackbone network to improve the detection effect of this model for small targets.The specific improvements are as follows:(1)In view of the low resolution of small target labeling frames,dense distribution and easy overlapping,this paper adds a deformable convolution module to the backbone network of YOLOv8.The added deformable convolution module can flexibly deal with the problem of insufficient receptive field of detection points corresponding to small targets,and strengthen the attention to small targets,so that the situation of missed detection and false detection is effectively improved,and the detection accuracy is improved.(2)In view of the problem that small target detection is susceptible to image background and noise interference,this paper adds an attention mechanism to the neck structure of YOLOv8.Using the characteristics of the attention mechanism to focus on more important key features,it can filter out relatively important information from a large amount of information,enhance the degree of attention to the underlying features,and strengthen the attention to small targets,thereby improving model detection.precision.(3)Aiming at the problem that the small target classification and localization loss is not easy to calculate,in the existing IoU loss function,this article propose a new IoU loss function calculation method in view of the ideas of wise-IoU.This calculation method can dynamically adjust the proportion of each part in the loss function at different stages of training,so that small objects can better return to the real label frame and improve the detection performance of small objects.From the experimental results we can conclude that our optimized model has better detection performance for small targets compared to YOLOv8s while maintaining training time and inference speed. |
