Research And Application On Fine-grained Image Classification Based On Regional Information Enhancement

Fine-grained image classification refers to the finer division of subcategories within a large category,such as whether a bird is a seagull or a goose.With the development of artificial intelligence,there is an increasing demand for sub-category classification of objects in the same basic category,such as brand classification of commodities,plant classification in the field of plant research,vehicle model and brand classification,etc.However,due to the small difference between fine-grained image categories and the large difference within the categories,fine-grained image classification is a very challenging task.Since subcategories usually have small inter-class differences,categories often need to be distinguished by small local differences.In the framework of deep learning,this paper studies how to enhance the attention to the local region,that is,how to enhance the regional information to improve the performance of fine-grained image classification network,and studies the application of this technology in vehicle retrieval and recognition.The main research results are as follows:(1)Aiming at the problem that hierarchical bilinear pooling(HBP)network affects the classification performance by carrying out feature interaction on all activated regions including irrelevant background.A saliency enhanced HBP(SE-HBP)network is proposed.Based on HBP,SE-HBP combines with the saliency detection network to generate an attention map,and uses the attention map to interact with the feature extraction network to enhance the information of the salient regions.As the result,it can reduce the impact of background and other irrelevant information.Finally,the classification accuracy of three commonly used fine-grained image datasets CUB-200-2011,Stanford Cars and FGVC-Aircraft is 86.5%,92.9% and 90.8%,respectively.(2)Aiming at the problem that the existing strong supervision methods rely excessively on extra manual annotations and have a large number of parameters.This paper proposes a component information distillation(CID)method for fine-grained image classification.In this method,the teacher net is first trained by using component labeling samples,then distills the component information and guides the student net to conduct training,so that the student net realizes the enhancement of component region information.The student net completes the training by receiving soft target provided by the teacher net.By using the student net test,only the original images are needed to obtain the high-precision results without adding additional parameters.The classification accuracy of CUB-200-2011 and Birdsnap is 88.0%and 81.3% respectively.(3)The vehicle retrieval system and identification system are built,and the application of CID method in vehicle retrieval and identification tasks is explored.First,the vehicle recognition model is trained by using the bayonet vehicle image and CID method,and then it is applied to construct the offline vehicle retrieval system and online vehicle identification simulation system.The vehicle identification network here finely classifies vehicles,that is,different models of vehicles of the same brand can be identified.Offline vehicle retrieval system can retrieve qualified vehicles from the bayonet vehicle database.Online vehicle identification simulation system is mainly aimed at road monitoring,which can recognize the vehicles in bayonet monitoring in real time.Finally,the graphical interfaces of offline retrieval system and online identification simulation system are constructed respectively,and the tests are carried out on the bayonet vehicle dataset and real road videos,which verify the validity of the research contents of this paper.