Fine-grained Vehicle Classification And Re-identification Based On Deep Learning

With the increasing number of vehicles in current transportation systems,intelli-gent video surveillance and management becomes more necessary which is one of the important artificial intelligence fields.Vehicle-related problems are being widely ex-plored and applied practically.Among various techniques,computer vision and deep learning algorithms have become the most popular ones,since a vast of video/image surveillance data are available for research,nowadays.In this paper,a wide range of researches has been conducted to solve the vision-based fine-grained vehicle classi-fication and vehicle re-identification problems.Moreover,many methods have been proposed to solve the problem,e.g.,complex application scenes,the insufficient com-puting capacity of terminal equipment,and low resolution of image and video data,and achieve state-of-the-art results.The main work and innovations of this thesis are summarized as follows:1.Vehicle plate recognition is important in vehicle re-identification system and has been explored even before the emerging of deep learning.Those methods,however,have disadvantages,e.g.,some frameworks have too many layers,which significantly slows the speed.To solve the problem,we proposed a light and efficient framework of plate recognition,which reformulates the task as recognizing digits,in contrast to many studies of recognizing the entire number plate.This method reduces the error of entire recognition effectively.The feature map in CNN architecture is cropped with different overlaps to realize a robust recognition.Moreover,the model is simplified to six layers for a promised speed.Experiments show that this method achieve both high accuracy and fast speed compared with all other methods.2.Researches of Fine-grained classification has been studying for many years and achieve excellent results.However,there is little research focusing on vehicle recog-nition specifically.To solve the problem,a novel architecture of Fine-grained vehicle classification is proposed.Firstly,a vehicle posture classification model is proposed to classify the vehicle to five postures.Secondly,a deep neural regression network is proposed to perform key-points detection.Based on detected key-points,key-regions are obtained.The feature of key-regions and whole vehicle are fused to realize an effi-cient feature embedding in a coarse-to-fine manner.Compared with the features from the global vehicle,the fused features are more descriptive and discriminant.Experi-ments show the method increases the accuracy of classification significantly,and meet the requirement of practical application.3.To address the problems of complex intra-class and inter-class variations of ve-hicle re-identification task,a refined part model with a novel vehicle re-identification network is proposed.Different from other methods,which directly obtain region part for vehicle re-ID,the refined part model is formulated through a Grid Spatial Trans-former Network to automatically locate the vehicle and perform division for regional features.Residual attention is also conducted to give an additional refinement for a fine-grained identification.Finally,the refined part features are fused to form an ef-ficient feature embedding.Experimental results show that our approach outperforms state-of-the-art.A hierarchical ranking loss is deployed to pull the images of the same vehicle together compactly and pull different vehicles as well as different vehicle mod-els apart.With the supervision of this ranking loss and refined part model,the network can learn a structured feature embedding space in a coarse-to-fine manner to enhance the intra-class compactness and inter-class distinction,thus characterizing multi-level semantic similarity between vehicle images.Experiments demonstrate the proposed method improves re-identification performance of state-of-the-art by about 6%.