YOLOV4-Balance:Object Detection Network Based on Sample Balance
In order to solve the problem that the algorithm has low detection accuracy for small targets due to insufficient number of small tar-get samples in the dataset and unbalanced distribution.Based on the multi-scale target detection model YOLOv4,YOLOv4-Balance was pro-posed which combines dynamic loss feedback and data augmentation.First,in order to balance the distribution of small target samples and en-rich the quality of small target samples in the data set,a data enhancement algorithm U-Mix based on image combination and stitching is pro-posed.Secondly,based on the Loss feedback during model training iterations,a multi-scale model training algorithm DLF(Dynamic loss feed-back)using dynamic Loss feedback is proposed to improve the contribution of the small target samples to the model during the training pro-cess.The experimental results show that in the MS COCO dataset,compared with the baseline model YOLOv4,the average accuracy of YO-LOv4-Balance is improved by 2.1％and the detection accuracy for small target samples is improved by 2.8％.The algorithm in this paper will not introduce additional computational overhead,and the model converges quickly,which is conducive to efficient training.
万方数据
维普
