Research On Recognition Method Of Heel-down Feature Of Three-dimensional Footprints Based On Morphological Examination

Analysis of gait characteristics of three-dimensional footprints is an important part of the field of criminal forensic science.The shape of three-dimensional footprints can reflect the personal characteristics of the people left behind by the footprints,which plays an important role in defining the scope of investigation and improving the efficiency of solving cases.The theoretical research on the analysis of three-dimensional footprint gait characteristics has been well inherited and developed in the years of public security work,and a complete theoretical system has been formed.However,it is inevitable that the analysis and identification of threedimensional footprint gait characteristics is subjective and empirical to a certain extent,which requires the knowledge and experience of evaluators.Moreover,it is difficult to guarantee the objectivity of identification,which limits the development of footprint research to a certain extent in today’s information age.In this paper,convolutional neural network is introduced into the recognition of heel-down features of three-dimensional footprint to realize the automatic recognition of heel-down features of three-dimensional footprints,and methods to improve the objectivity and intelligence of footprint inspection are discussed and tried.In this paper,the convolutional neural network is used to study the automatic recognition of four kinds of gait features,namely,the tread mark,the scratch mark,the knock mark and the push mark.In this experiment,the three-dimensional footprint data set was first established.In the experiment,26 pairs of experimental shoes including 13 pairs of leather shoes and 13 pairs of training shoes,basketball shoes,running shoes and other shoes were collected,among which20 pairs of shoes were training set experimental shoes and 6 pairs of shoes were test set experimental shoes.13 volunteers were organized.Among them,10 volunteers collected the images of the training set and 3 volunteers collected the images of the test set.They were asked to wear these experimental shoes and leave stereoscopic footprints on the sand respectively.A total of 120,368 three-dimensional footprints were obtained,including 100,512 images of the training set and 19,856 images of the test set.Le Net model,Alex Net model and VGG-16 model were used for feature extraction and recognition experiments,and the recognition accuracy rates were 83.56%,90.10% and 86.78%,respectively.Thus,it can be seen that both Lenet with 2convolution layers and VGG-16 with 13 convolution layers are not suitable for the recognition of heel-down feature,while Alex Net with 5 convolution layers is more suitable.Therefore,for the recognition of heel-down feature of three-dimensional footprints,the number of convolutional layers should not be too many or too few,and about 5 layers is more appropriate.Secondly,in order to further improve the accuracy of recognition,Footprint Net,a new network model,is proposed based on the above experimental results,which has a good recognition effect on the three-dimensional footprint data set.After that,an appropriate normalization method and activation function were selected to improve the recognition accuracy to 92.13%.Visualized feature images of three-dimensional footprint images show that the network model can effectively extract the heel-down features from three-dimensional footprint images.Finally,in order to solve the problem of gait feature recognition of fuzzy and deformed three-dimensional footprint images in actual public security work,the second data set,namely low-quality three-dimensional footprint image data set,was established in this experiment.This data set is based on the three-dimensional footprint data set and uses image affine transformation to imitate the low-quality three-dimensional footprint images in other photos or videos of the scene in five directions and rotate them in three directions.There are 192,400 images in total,including 160,400 images from the training set and 32,000 images from the test set.For this data set,Footprint Net was used to carry out the identification experiment of heeldown features,and the identification accuracy was 89.12%.Then the Residual structure was added,which improved the recognition accuracy to 93.48%.Thus,it can be seen that the optimized Footprint Net model can achieve better recognition results for the heel-down feature of low-quality three-dimensional footprint images.This study verified the feasibility of using convolutional neural network to automatically recognize the heel-down feature of three-dimensional footprint,and realized the recognition of the heel-down features in high-quality images and low-quality images with high accuracy,which is a beneficial exploration for the automated footprint inspection today.