基于微调VGG-16的现场鞋印检索算法

Research on Forensic Shoeprint Retrieval Algorithm by Fine-tuning VGG-16

鞋印是刑事案件中出现率最高的痕迹物证[1],通过检索现场鞋印进而确定鞋样能够获得有关嫌疑人身份和犯罪特点的重要信息。近年来研究人员逐渐把深度学习的相关方法应用到鞋印检索上,但目前大多数基于深度学习的鞋印检索算法都直接使用预训练的卷积神经网络提取特征,并未微调再训练,也没有设计并训练新的网络模型。提出一种基于微调VGG-16的现场鞋印检索算法。首先建立一个432类共2827幅图片的鞋印数据集,并进一步增广到228987幅图像。然后使用该数据集微调ILSVRC数据集预训练的VGG-16模型,并将该模型作为鞋印特征提取器。实验结果显示,与使用预训练模型相比本文方法的检索精度有了明显提高,在200幅嫌疑鞋印和5000幅样本鞋印图像构成的测试数据集上top10的正确识别率达75.5%。

Shoeprints are the most common trace evidence in criminal cases[1].By searching the shoeprints in the database,we can determine the shoe pattern and further important information about the identity of the suspect and the characteristics of the crime can be obtained.In recent years,although researchers have gradually applied the relevant methods of deep learning to shoeprint retrieval,most of the shoeprint retrieval algorithms based on deep learning directly used pretrained convolutional neural network to extract features without fine-tuning or designing and training new models.This paper proposes a forensic shoeprint retrieval algorithm based on VGG-16.First,we establish a shoeprint dataset with 2827 images in 432 categories and the data was further expanded to 228987 images.Second,we retrain the VGG-16 model trained by ILSVRC dataset using our data and use the retrained model as a shoeprint feature extractor.The experiment result shows that the accuracy has obvious improvement compared with the pre-trained model.The identification rate of our method is 75.5%at top 10 on the test dataset containing 200 probe shoeprints and 5000 refer shoeprints.