Research On Automatic Identification Method Of Hair Evidence Micrographs Based On Morphological Examination

Hair,as one of the common evidence at crime scenes,is of great significance for crime scene reconstruction analysis,determination of investigation direction,exclusion of criminal suspects and trajectory tracking.At present,the morphological examination method of hair evidence mainly uses microscope to observe the micro-morphology and complete taxonomic identification,which is mature but low in automation,prone to subjective errors.This paper proposes a hair micrograph classification method based on convolutional neural network for hair evidence identification.The self-built network is used to extract and train the morphological features of micrographs to realize the automatic classification of hair evidence and improve the intelligence level of on-site evidence inspection.In this paper,a Leica DVM6 digital microscope equipped with LAS X software is used to collect micrographs of different types of pre-made hair sample slices before image processing,such as sliding cropping,grayscale conversion,and data enhancement.The Human Hair Micrograph Experimental Data Set was established,which contained 5 types and a total of 30,000 images.And the Human & Animal Hair Micrograph Experimental Data Set was established,which contained 10 types and a total of 300,000 images.According to the characteristics of two data sets,Hair-Net is designed to conduct sample training and testing on the Human Hair Micrograph Experimental Data Set,and constantly adjust parameters while optimizing structure to obtain the best model Hair-Net-Human.Then use the improved network to perform sample training and testing on the Human & Animal Hair Micrograph Experimental Data Set,comparing with classic networks LeNet and AlexNet.The network optimization method is introduced to gradually improve the network performance to obtain the best classification model Hair-Net-Animal.Finally,the practical application effect of the two-level automatic classification mode and best models is tested by technical blind test.The experimental research shows that: 1.The designed convolutional neural network Hair-Net has a good classification effect on the Human Hair Micrograph Experimental Data Set.Adjust model parameters such as the learning rate,number of iterations and the batchsize,then optimize the structure of input layer while introducing the Leaky ReLU activation function and Batch Normalization.The best model Hair-Net-Human with great generalization ability was finally obtained,and classification accuracy of which is up to 96.93%.2.The improved Hair-Net model is better than classic networks LeNet and AlexNet in classification of the Human & Animal Hair Micrograph Experimental Data Set.While using Inception structure to lightweight the network,Center-loss function is introduced for joint supervision with original Softmax-loss function,which can significantly improves network performance and robustness,the classification accuracy of the best model Hair-Net-Animal can reach 95.24%;3.Design a simulation site to extract hair evidence and collect micrographs.According to the two-level automatic classification mode of hair micrographs established in this paper,input the extracted images step-by-step into hair micrograph classification model Hair-Net-Animal and Hair-Net-Human for classification detection,the overall accuracy are both above 80%,which verify the feasibility of the method in practical application.This article breaks the traditional identification method of hair evidence,creatively combines deep learning,designs and builds the new network Hair-Net,which has achieved a good classification effect on the self-built hair micrograph data sets.A two-level automatic classification mode of hair micrographs is established.then simulation test is performed to verify that the method in this paper can achieve rapid automatic classification after hair evidence collection and effectively reduce subjective errors of manual identification,which provides a new automatic classification method for hair morphology inspection.