为了提高行人属性识别的准确率,提出了一种基于多尺度注意力网络的行人属性识别算法。为了提高算法的特征表达能力和属性判别能力,首先,在残差网络ResNet50的基础上,增加了自顶向下的特征金字塔和注意力模块,自顶向下的特征金字塔由自底向上提取的视觉特征构建;然后,融合特征金字塔中不同尺度的特征,为每层特征的通道注意力赋予不同的权重。最后,改进了模型损失函数以减弱数据不平衡对属性识别率的影响。在RAP和PA-100K数据集上的实验结果表明,与现有算法相比,本算法对行人属性识别的平均精度、准确度、F1性能更好。

In order to improve the accuracy of pedestrian attribute recognition, a multi-scale attention network for pedestrian attribute recognition algorithm is proposed in this paper. In order to improve the ability of feature expression and attribute recognition of the algorithm, first, the top-down feature pyramid and attention module are added to the residual network ResNet50. A top-down feature pyramid is constructed from the visual features extracted from the bottom-up. Then, the features of different scales in the feature pyramid are fused to give different weights to the channel attention of each layer of features. Finally, the model loss function is improved to weaken the impact of data imbalance on the attribute recognition rate. Experimental results on the RAP and PA-100K data sets show that compared with existing algorithms, the algorithm has better performance in terms of average accuracy, accuracy, and F1 for pedestrian attribute recognition.