Research On Bullet Marks Recognition Based On Transfer Learning

With the deeper progress of the crime crackdown,fighting against gun violence has become one of the key tasks of the current public security work.The inspection of gunshot traces provides important evidence support for the detection of gun-related cases,and is the core technology to combat gun-related crimes.The traditional bullet mark detection technology uses the comparison microscope to examine the surface of bullet mark through 2D image comparison,line butt joint and other methods.This method has the characteristics of finding defects such as low efficiency,high work intensity and no automatic batch comparison.It has the disadvantages of hard to find features,low comparison efficiency,high work intensity,and inability to automatically compare in batches,etc.Currently,a 3D topography measurement using tiny marks is used in the bullet mark detection.It reproduce three-dimensional characteristics of bullet marks by the high resolution 3D point cloud data and the method of 2D image composite.It could complete the digital archive of bullet mark,and identify bullet mark feature based on 3D data and 2D image.However,the current method based on 3D topography data,has significantly improved merely in bullet mark features identification,speed and efficiency.For multiple bullet mark evidence sample and massive bullet sample archive recognition,need an entirely new approach,which can find the match with on-site inspection of bullet mark in the archive.In this way,it can quickly identify suspected guns and lock down criminal suspects.With the characteristics of local perception and parameter sharing,convolutional neural network can effectively learn the corresponding features from samples and avoid the complex feature extraction process by using the original image as input.At present,convolutional neural network has been widely used in image classification,speech recognition and target recognition.In this thesis,a method of bullet marks identification based on convolutional neural network is proposed.The main research contents and works are as follows.(1)Three main traditional methods of bullet mark identification are analyzed.Two-dimensional image comparison method,section depth curve comparison method and three-dimensional reconstruction comparison method.(2)A bullet marks identification method combining transfer learning and convolutional neural network was proposed.The convolutional neural network is used to improve the efficiency and accuracy of identification.And the FaceNet is transferred to solve the problem that using convolutional neural network requires a large number of data sets.The main steps of this method include: first,preprocess the depth data and process the data into the input form conforming to the FaceNet model;Secondly,the pre-training model of FaceNet is loaded to obtain the feature vectors of depth data;Finally,using a method similar to online learning,adaptively adjust the threshold to get the final bullets similarity ordering.The theoretical analysis and experimental results show that compared with the traditional method of curve similarity comparison based on the depth data of the bullet marks,this method is more accuracy and efficiency.(3)The harmony search algorithm is used to optimize the weight of FaceNet.It solves the problem that when using “Triplet loss” in the training,the appropriate Triplet needs to be specially selected,otherwise it will difficult to converge.The experimental results of identification accuracy,F1-score and ROC curve before and after the optimization show that using the harmony search algorithm to optimize FaceNet weight has a good optimization effect in bullet marks identification.