基于smoothL1改进的边框回归损失函数

Improved Bounding Box Regression Loss Function Based on SmoothL1

为了进一步提高目标检测任务中的边框回归精度,提出一种基于smoothL1改进的边框回归损失函数.通过自适应地增大smoothL1中非离散点的梯度,缓解了模型反向传播中对离散点和非离散点梯度分布不平衡问题,提高了模型的精度.实验结果表明,在PASCAL VOC2007测试集上,基于改进的smoothL1的目标检测模型Faster R-CNN,平均精度均值(mAP)达到了70.8%,相较smoothL1,模型精度有所提高.

In order to improve the accuracy of the bounding box regression in object detection,an improved bounding box regression loss function based on smoothL1 is proposed.By adaptively improving the gradient of the non-discrete points in smoothL1,the problem of the imbalance of the model’s gradient distribution to discrete points and non-discrete points in the back propagation is alleviated,and the accuracy of the model is improved.Experimental results show that on PASCAL VOC 2007 test dateset,based on the improved smoothL1 object detection model Faster R-CNN,and the mean average accuracy(mAP)reaches 70.8%.Compared with smoothL1,the model accuracy is improved.