Study On Fine-grained Image Classification Method Based On Dictionary Learning

Image classification has become a hot topic in the field of computer vision,due to its wide range of practical application value.The collaborative representation based classification algorithm can solve the general image classification problem very well.However,it is difficult for the traditional image classification algorithm to solve the fine-grained image classification problem,which has big difference within the class and small scatter among the classes.To deal with the problem,experts and scholars in related fields carried out a series of studies from two aspects,which are the amelioration of feature extraction methods and the improvement of multi-class classification algorithms.In this thesis,we extract the image features via the convolutional neural networks first.Then we propose class specific dictionary learning algorithms to improve the performance of the collaborative representation based method for fine-grained image classification.Finally,we evaluate our proposed algorithms on several benchmark fine-grained image datasets.The main work is threefold:1.The structure and characteristic of convolutional neural networks are studied and the image features are extracted via convolutional neural networks.To compensate the lack of spatial information by convolutional neural networks,the spatial pyramid matching algorithm is utilized during the procedure of image feature extraction.2.A class specific dictionary learning based kernel collaborative representation algorithm is proposed.Based on the collaborative representation based classification algorithm,the process of dictionary training is added to assign different weights to the training samples.Furthermore,the proposed algorithm is extended to the kernel space to obtain the nonlinear information hidden in the image features.The experimental results on five datasets show that the class specific dictionary learning based kernel collaborative representation algorithm improves the image classification performance to a certain extent.3.A class specific centralized dictionary learning based kernel collaborative representation algorithm is proposed.On the basis of the class specific dictionary learning based kernel collaborative representation algorithm,within class difference is considered as a regularization constraint to make the sparse coding of the same class more concentrated and reduce the divergence within class.The experimental results on five datasets show that the class specific centralized dictionary learning based kernel collaborative representation algorithm improves the image classification performance obviously.