Study On Feature Extraction For Tire Tread Pattern Image Classification

With the proposal of “safe city” planning for national construction,traffic safety and city public security management have become particularly critical.In criminal scene investigation and traffic accident handling,tire tread pattern comparison can provide useful clues for police to lock the suspect vehicle and divide accident liability.With the development of economy,the number of cars in China has been increasing rapidly,the quantity of tread pattern images collected by police on the scene has exploded.Therefore,study of tread pattern classification algorithm,using information technology to automatically classify and manage tread pattern images,to help police in improving their working efficiency.It's a very challenging research topic in the work of strengthening police by science and technology in our country,it has important application value for public security organizations to maintain social order.The key technology of tread pattern classification is to design image feature extraction algorithm which can effectively describe tread patterns.Relying on the Key Laboratory of Electronic Information Application Technology for Scene Investigation,co-built by Xi'an University of Posts and Telecommunications and the Department of Public Security of Shaanxi Province.According to the application requirements of public security criminal investigation industry,combined with the characteristics of tread pattern images,two kinds of image feature extraction algorithms suitable for tire tread pattern classification are proposed.These two algorithms effectively improve classification accuracy of tire tread patterns.The main contributions of the thesis are as follows,Firstly,with the support of the key laboratory platform,the author worked with other team members to collect tread pattern images at different times,in different locations and under different shooting conditions,built a tread pattern image dataset for academic research,which plays a positive role in promoting the research and development in this field.Secondly,through the systematic investigation of tread pattern image classification field,the research status in this area is described and summarized in detail.Tread pattern image dataset used in literature are introduced and performance evaluation parameters used for TPIC are summarized.The research achievements are discussed and analyzed based on two key technologies: low-level feature extraction and high-level semantic analysis.In addition,according to application requirements in public security area,the research challenges in this field are analyzed and future research trends are pointed out.Thirdly,leveraging on the directionality characteristics of tread patterns,a novel tread pattern feature extraction algorithm is proposed based on adaptive weighted feature fusion with the weights defined by sub-band energy ratio.The proposed approach consists of:(1)discrete wavelet decomposition of tire tread image to obtain low frequency,horizontal,vertical and diagonal sub-bands;(2)extraction of improved local binary pattern features from the sub-band images;(3)concatenating the tread pattern directional features,weighted by their corresponding sub-band energies,then applying SVM for tire tread pattern classification.Then,to relieve the over-fitting problem in CNN model training due to the lack of large scale training data,an effective tread pattern image feature extraction algorithm based on transfer learning is proposed.Transfer learning is introduced into model training.The parameters of a pre-trained model are fine-tuned using tread pattern image data,and a new model for the task of tread pattern classification is produced.The fully connected layer features are extracted from the new model as the high-level features and weighted with the low-level features proposed in chapter 3 as fusion features to train MKSVM for tread pattern image classification purpose.Finally,in order to facilitate academic communication and achievements presentation,a tire tread pattern image classification demonstration system is designed.The demonstration system can realize tread pattern images input,image feature extraction,SVM classifier training and classification.The demonstration system is simple and quick,easy to operate,the interface is friendly and interactive,which is convenient for users.In summary,this thesis reviews the field of tread pattern image classification,and proposes two image feature extraction algorithms suitable for tread pattern classification.The related research results provide a new scheme and a new idea for solving the problem of tread pattern classification,and have certain reference value for the related research.