Target detection method based on improved YOLOv3
YiXin Jia, Luyang Liu, Lijie Zhang, Hai Li · 2022 3rd International Conference on Computer Vision, Image and Deep Learning & International Conference on Computer Engineering and Applications (CVIDL & ICCEA) · 2022
With the continuous development of the field of computer vision and the need for the accuracy and timeliness of model recognition in practical applications, YOLOv3 draws on YOLOvl and YOLOv2. Although there are not many innovations, it maintains the speed advantage of the YOLO family and the detection accuracy is improved, especially for the detection ability of small objects [1]. However, in the three-layer feature network structure of YOLOv3, the correlation between the layers is not strong, which may lead to the situation that the information of the input image cannot be used efficiently, thus affecting the final accuracy. In this regard, this paper mainly studies the improvement method for the YOLOv3 feature pyramid network, and there are many improvement methods for this network. This paper uses residual linking and attention mechanism methods. The fusion judgment of features is realized by introducing the CBAM attention module and adding residual links in the same layer and residual links across layers. Finally, we use the VOC07+12 dataset [2] to obtain an accuracy of 79.5% under the condition of IoU=0.75 after 150 iterations, which is an improvement compared to the 73.6% accuracy of the original model.