基于鞋印图像的性别预测方法

鞋印是作案人在案发现场经常遗留的痕迹，承载人的性别、身高等属性信息。基于鞋印的性别预测对快速排查嫌疑人具有重要作用，其方法主要由刑侦人员凭借经验判断，需要大量领域知识，而少数自动预测方法是基于人工提取的特征和经验模型进行预测，受测量误差的影响，导致预测准确率降低。针对该问题，提出基于鞋印图像的端到端预测方法。采用卷积神经网络提取鞋印图像特征，引入通道注意力模块对特征权重进行重新分配，使模型重点关注鞋印图像中对性别起显著作用的部分。在此基础上，将特征图输入到性别预测模块进行预测。此外，分别构建适用于单枚和多枚鞋印应用场景的数据集SiSIS和SeSIS，根据在案发现场中鞋印可能出现的情况，设计鞋印方向差异、鞋印残缺和弹性形变的数据增广方式。实验结果表明，该方法在SiSIS和SeSIS数据集上的预测准确率分别达到91.80%和99.35%，相比现有基于鞋印的性别预测方法，具有较优的预测性能。

Gender Prediction Method Based on Shoeprint Image

Shoeprints are traces frequently left behind at crime scenes.They provide information regarding a person's attributes, such as gender and height.Gender prediction based on shoeprint plays a critical role in the rapid screening of suspects.Currently, shoeprint-based gender prediction method is mainly based on the experience of criminal investigators, which depends on extensive domain knowledge.Although a few automatic prediction methods are based on manually extracted features and empirical models, they are affected by measurement errors, reducing the prediction accuracy.This study proposes an automatic end-to-end shoeprint image-based gender prediction method.The convolutional neural network is used to extract shoeprint image features.The channel attention module is introduced to redistribute the feature weights such that the model focuses on the parts of the shoeprint image that play a significant role in gender prediction.The feature image is input into the gender prediction module for prediction using the proposed model.In addition, SiSIS and SeSIS datasets suitable for single and multiple shoeprint application scenarios are constructed.Based on possible cases of shoeprints at the crime scene, data augmentation methods of shoeprint direction difference, shoeprint imperfection, and elastic deformation are designed.The experimental results show that the prediction accuracy of the proposed method on SiSIS and SeSIS datasets are 91.80% and 99.35%, respectively.The proposed method performs better than existing gender prediction methods based on shoeprints.