基于改进损失函数的多阶段行人属性识别方法

Multistage Pedestrian Attribute Recognition Method Based on Improved Loss Function

大量研究工作通过挖掘属性间的正相关性提高视频监控场景下的行人属性识别性能,但对属性间负相关性的探索仍存在不足.为此,文中基于深度学习理论提出多阶段行人属性识别方法,同时探索属性间的正、负相关性.第一阶段计算每个属性在训练过程中的损失值和正确率.第二阶段为平均损失较大且正确率较小的属性单独建立一个网络分支,其它属性仍保留在原分支上,然后两个分支联合预测所有属性.第三阶段新建两个网络分支,结构与第二阶段的分支相同,优化新分支的参数,使其属性识别性能优于第二阶段.最终使用第三阶段的模型进行属性预测.此外,构建增大正负样本差异的改进损失函数,应用于三个阶段的训练,进一步提升模型性能.在两个行人属性识别数据集RAP和PETA上的实验表明,文中方法性能较优.

There are plenty of studies on improving the performance of pedestrian attribute recognition in video surveillance scenarios by mining the positive correlations between attributes. However, the research on negative correlations is far from enough. In this paper, a deep learning based multi-stage pedestrian attribute recognition method is proposed to simultaneously explore the positive and negative correlations between attributes. In the first stage, the loss value and the accuracy of each attribute are calculated during training. In the second stage, a new network branch is designed for the attributes with larger average loss and lower average accuracy, while other attributes remain on the original branch. All attributes are predicted by these two branches jointly. In the third stage, two new network branches with same structure as the second stage are designed to optimize the parameters and improve the performance during attribute recognition. Moreover, an improved loss function increasing the distance between positive and negative samples is proposed, and it is applied in all training stages to further improve the performance. Experiments on datasets RAP and PETA validate the promising performance of the proposed method.