基于空洞卷积和Focal Loss的改进YOLOv3算法

Improved YOLOv3 based on dilated convolution and Focal Loss

为了进一步提升YOLOv3的小目标检测能力,文中提出将Darknet-53中的第2个残差块输出的特征图用混合空洞卷积处理后,与YOLOv3的8倍下采样特征图相融合,建立新的检测特征;同时,使用Focal Loss改进损失函数中的负样本置信度公式,缓解YOLOv3的正负样本比例失衡问题。实验结果表明,在小目标数量占比为47.7%的特定测试集上,改进YOLOv3的平均准确率和召回率分别比原YOLOv3提高了8.8%和16%;在VOC测试集上,改进YOLOv3的平均精度均值比原YOLOv3提升了3.4%。

To improve the detection performance of small targets of the YOLOv3 algorithm, a technique for producing new prediction features is proposed.It fuses the output of the hybrid dilated convolution on the second residual block in Darknet-53 and the output of the 8 x down-sampling of the original network.In addition, the Focal Loss is used to compute the confidence of negative samples of the loss function, alleviating the imbalance problems of positive and negative samples of YOLOv3.Experimental results show that compared with the YOLOv3 algorithm, the improved YOLOv3 algorithm can achieve an improvement of 8.8% and 16% in terms of the average precision and the recall on a specific data set with a proportion of the number of small targets of 47.7%,and achieve an improvement of 3.4% in terms of the mean average precision on the VOC dataset.