Research And Application On Fine-Grained Image Classification Based On Bilinear Model

Fine-grained image classification is a more refined classification task aims to classify hundreds of subcategories which belong to the same basic-level category,while fine-grained level images are the images with more refined subcategories.With the rapid expansion of the image scale and the increasing demand of human for classification and retrieval,the traditional basic-level image classification become increasingly unable to meet the needs of recognition tasks because of its low classification accuracy.Fine-grained image classification has gradually replaced basic-level image classification as a hot research topic in the field of computer vision.B-CNN model has attracted wide attention in fine-grained image analysis due to its simple network architecture and high classification accuracy.This paper studies the fine-grained classification problem based on B-CNN from some aspects:subcategory similarity measurement,multi-layer convolution feature fusion,et al.The main research contents are as follows:(1)Aiming at the problem that different subcategories are hard to be distinguish due to the effect of their backgrounds,a bilinear fine-grained classification method based on weakly supervised localization and subcategory similarity measurement is proposed.By selecting the convolution descriptors,the model can eliminate the influence of background noise and obtain more accurate features,a fuzzy similarity matrix is then obtained to measure the similarity among the subcategories according to the results of the training set.The fuzzy similarity matrix is used to modify the loss function,which makes it easier to distinguish different subcategories.Extensive experiments implemented on Stanford Cars-196 and CUB-200-2011 show that the proposed method can effectively improve the accuracy of classification.(2)Although weakly supervised localization and similarity measurement eliminate the interference of background noise to a certain extent,enhancing the ability to classify confusing categories.Since the differences between fine-grained images mainly exist in small local areas,while most existing algorithms only extract the features of last convolution layer and ignore local information,a multi-layer feature fusion method based on B-CNN is proposed.The bilinear features between and within different convolution layers are cascaded to obtain the final feature,and softmax is used to classify.In order to reduce the number of parameters,the parametric matrix is decomposed into the product of two first-order vectors by decomposition bilinear pooling to avoid high-dimensional matrix operations.Experiments on CUB-200-2011 and Stanford Cars-196 show that the proposed method can effectively fuse multi-layer features and improve the accuracy of fine-grained image classification.(3)In order to achieve the effect of fast image retrieval,based on the above research results,a fine-grained image retrieval system based on B-CNN is designed.The system mainly consists of three modules:image database building,model training and fine-grained image retrieval.Particularly,Manhattan distance is used to measure the similarity between retrieved images and the images in the library.Practical application shows that the system can effectively retrieve fine-grained images.