基于卷积神经网络的毛发显微图像分类

Classification of hair microscopic image based on convolutional neural networks

显微照相在现场毛发物证提取与采集中被广泛应用,而基于显微形态图像处理分析的毛发识别有助于进一步提高毛发检验鉴定的自动化程度。为实现毛发的无损检验,首先利用光学显微镜进行不同来源、不同条件的毛发形态显微图像的采集,使用Matlab软件预处理400倍显微镜下拍摄的样本图像数据;尝试多种卷积神经网络的结构和参数组合对毛发显微图像进行特征提取与训练学习,加入中心损失函数提升识别率及泛化能力来优化网络,将显微图像输入网络即可实现毛发的快速识别。本文对人类毛发、假发等四类共20 000张显微图像进行分类训练及测试实验。结果表明,在学习率为0. 000 7,训练迭代次数为30 000次时,识别率达98. 97%,泛化精度达98. 80%。该方法可实现毛发显微图像的高效、准确、自动分类识别,可提高毛发鉴定效率。

Photomicrography is widely used in the field of hair material evidence extraction and collection,and hair recognition based on microscopic image processing analysis helps to further improve the automation of hair inspection and identification.In order to achieve non-destructive testing of hair,firstly,optical microscopy was used to collect hair microscopic images from different sources and under different conditions,and sample image data taken under 400 times microscope was preprocessed using Matlab software.Extract and train the features of hair microscopy image combined with various structures and parameters groups of convolutional neural networks the parameters.Add the central loss function to improve the recognition rate and generalization ability to optimize the network,and input the microscopic image into the network to realize the rapid recognition of hair.In this paper,a total of 20,000 microscopic images of human hair,wig and other four types of classification were used in training and testing experiments.The results show that when the learning rate is 0.000 7 and the number of training iterations is 30,000,the recognition rate is 98.97%,and the generalization accuracy is 98.80%.The method can realize efficient,accurate and automatic classification and recognition of hair microscopic images,and can improve hair identification efficiency.