根据现场遗留鞋印推断出作案人所穿鞋型，再从周围监控视频中搜索嫌疑鞋型已成为公安机关侦破案件的重要技战法。针对该技战法完全依赖人工筛查、受主观影响大、易造成漏检等问题，提出了一种基于注意力机制的鞋型识别算法。首先，建立了贴近公安刑侦实战、样本容量为300的多背景监控鞋型数据集。然后，提出了一种注意力机制模型，以增强残差网络（ResNet50）对鞋子重要特征的提取能力。最后，对比了选取不同特征层输出作为鞋子特征及不同卷积特征聚合方法对识别精度的影响。为了增强模型的泛化能力，在损失函数中加入Label Smoothing。在多背景数据集上的实验结果表明，本算法的Rank-1、平均精度均值分别达到74.32%和56.97%。

It has become an important technique and tactics for the public security organs to infer the type of shoes worn by the perpetrators according to the shoe prints left at the scene, and then search the suspected type of shoes in the surrounding surveillance video. This technique is completely dependent on manual screening, which is greatly affected by subjective factors and easily leads to problems such as missed detection. To solve this problem, this paper proposes a shoe type recognition algorithm based on attention mechanism. First, close to the actual combat of public security criminal investigation, a multi background monitoring shoe data set with sample size of 300 is established. Then, an attention mechanism model is proposed to enhance the ability of the residual network (ResNet50) to extract important features of shoes. Finally, the effects of selecting the output of different feature layers as shoe features and different convolution feature aggregation methods on the recognition accuracy are compared. In order to enhance the generalization ability of the model, label smoothing is added to the loss function. The experimental results on the multi background data set show that the Rank-1 and mean average precision of the algorithm are 74.32% and 56.97%, respectively.