基于卷积神经网络的现场勘查照片分类方法

Method for Classifying Crime Scene Photographs Based on Convolution Neural Network

随着人工智能技术的迅速发展与广泛应用,智能化勘查方法正成为刑事科学技术领域新的研究热点,而实现现场勘查照片自动识别与分类是智能化勘查的重要研究内容。面向公安机关实战应用需求,提出了一种基于卷积神经网络的现场勘查照片自动分类算法。基于真实案件照片,建立了现场勘查照片数据集,包含现场勘查照片13164张,负类照片4008张。根据现场勘查照片数据特性,设计了现场勘查照片分类网络(CriSNet),通过对卷积层增加归一化处理以及改进bottleneck模块,实现对现场勘查照片的精确分类。实验结果表明:CriSNet模型的分类精度优于基准网络1个百分点,具有较好的鲁棒性,同时在分辨率低、品质较差的情况下,仍能保持较好的分类性能。

With the rapid development and wide application of artificial intelligence,intelligent investigation is becoming a new research hotspot in forensic science,and the realization of automatic recognition and classification of crime scene photographs is an essential aspect of intelligent investigations.We present an algorithm that automatically classifies crime scene photographs based on a convolution neural network.First,based on the data from criminal cases,a crime scene photograph dataset was constructed comprising 13164 scene photographs and 4008 negative photographs.Second,crime scene photograph net(CriSNet)was designed based on the data characteristics to accurately classify crime scene photographs by adding normalization processing to the convolution layer and improving the bottleneck module.The experimental results show that the accuracy of CriSNet is 1 percentage point better than that of the benchmark with good robustness,and CriSNet can still maintain excellent performance under low resolution and poorquality conditions.