Research And Application On The Algorithm For Fine-grained Vehicle Image Analysis

Computer vision is the process of using machines to understand and analyze image,in which fine-grained image analysis(Fine-grained Image Analysis,FGIA)has always been a hot research topic.The small inter-class variations and the large intra-class variations caused by the fine-grained nature makes it a challenging problem.In recent years,the technology of vehicle image analysis has been widely applied in the field of intelligent transportation,among which fine-grained image analysis method has been found to be effective in solving vehicle retrieval and vehicle identification problems.Compared with traditional image recognition and content-based image retrieval,fine-grained image analysis based on deep learning achieves higher performance and accuracy.This dissertation describes the research progress of fine-grained image analysis based on recognition and retrieval,and systematically introduces the relevant theories and methods used in this field.Aiming at vehicle image analysis,this dissertation proposes a more efficient fine-grained vehicle image recognition and retrieval framework based on previous research in the field of fine-grained image analysis,which not only effectively improves the accuracy of analysis,but also can be used as a plug and play unit.In addition,the research results presented in this dissertation are applied to intelligent vehicle retrieval system,which realizes the identification and retrieval of bayonet-electronic police image and surveillance video image.Finally,we discuss other application scenarios and future research directions of fine-grained image analysis.In this dissertation,a series of studies are carried out on fine-grained vehicle image analysis,the main work includes the following aspects:1.In fine-grained vehicle identification,we propose a Semantic Interaction Learning Network(SIL-Net).Instead of learning features of discriminative regional from an individual image,it is inspired by the twin network structure,aiming to discover semantic differences between two fine-grained categories through pair-comparison.Specifically,SIL-Net firstly collects contrastive information by learning mutual feature of input image pair,and then compares it with individual feature to generate corresponding semantic features.These semantic features learn differences from contextual comparison,this gives SIL-Net the ability to distinguish between two confusing images via interactive attention.After training,SIL-Net can adaptively learn feature priorities under the supervision of Margin Ranking Loss and converge quickly.SIL-Net achieves 94.8% and 98.5% accuracy in two public vehicle benchmarks(Stanford Cars and Comp Cars)respectively.2.In fine-grained vehicle retrieval,Recent studies have shown that ensembling different models and combining multiple global descriptors lead to performance improvement.However,training different models for the ensemble is not only difficult but also inefficient with respect to time and memory.This dissertation proposes a novel framework that exploits multiple global descriptors to get an ensemble effect while it can be trained in an end-to-end manner.The proposed framework is flexible and expandable by the global descriptor,Convolutional Neural Networks(CNN)backbone and weighted-aware moderate triplet loss.Moreover,we investigate the effectiveness of combining multiple global descriptors with quantitative and qualitative analysis.In benchmark tests,the proposed framework showed 98.9% and 99.1% retrieval accuracy on Stanford Cars and Comp Cars.3.The Semantic Interaction Learning Network for fine-grained vehicle recognition and the Diversity Attribute Integrationthe framework of fine-grained vehicle retrieval presented in this dissertation are applied to the traffic-video-image based vehicle retrieval system,realizing the recognition and retrieval functions of vehicle images respectively.Specifically,the user can extract the target vehicle image from video stream,or can enter bayonet or electronic police image.Meanwhile,the system can provide specific information of the target vehicle through the vehicle image recognition function,such as vehicle model,vehicle color and so on.Users can import the retrieval result image to the airborne library,such as the retrieval results taken by the sorting,so that users can acquire the vehicle’s key clues of information clearly,such as the vehicle’s trajectory,this information can help the traffic police and other related personnel to quickly find illegal,or criminal suspects vehicles,to a certain extent,it greatly improves the efficiency of intelligent traffic security management.