Fine-grained Classification Algorithm Research Based On B-CNN Model

In recent years,with the growth of big data,the task of image classification has been developed rapidly.The classification task has also evolved from two categories to simple coarse classification and then developed into fine-grained classification.Here the coarse classification of most cases is to distinguish between different objects,such as the cat,dog,tree,car,there is a significant difference between them.Fine-grained classification is to subdivide the same object,such as different types of birds,different models of aircraft,face recognition,etc.The difference between them is usually very small.The angles,light make the classification task more difficult.With the development of deep learning,image classification task becomes more and more easy,people's requirements for image classification is also increasing,fine-grained classification task came into being.Such as the variety of flowers,the type of birds,face recognition.The image classification of machine learning consists of two parts: images' feature extraction,and image classification with the features.Feature extraction is directly related to the classification accuracy,the greater the feature dimension,the higher the accuracy rate.However,when the feature dimension is too high,it will lead to more computer memory and the bigger amount of calculation.Convolution neural network is often used to extract features.Generally speaking,the deeper the network,the better the features,and the classification accuracy is also increasing.But when it comes to a certain depth,the accuracy becomes more and more difficult to improve by increasing the depth.So the researchers have proposed a number of methods to solve this problem,such as region based model,but its drawback is low efficiency,manual marking task is heavy.The B-CNN model solves these two problems,it needs only the labels of the train samples,and keep a higher accuracy rate.The B-CNN model solves the problem of fine-grained classification task to a certain extent,but its training and classification is for all samples.So it is difficult to distinguish that confusing classes.In this thesis,we propose a hierarchical B-CNN model guided by classification error.First,we propose the clustering algorithm guided by classification error to obtain the clusters in which the categories are misclassified to each other frequently.This algorithm is based on the constrained Laplacian rank methods.Through comparing the true labels and the classification results we get the ‘classification error matrix'.Then we construct the ‘affinity matrix' which will be used in the clustering process.Second,we propose the new hierarchical B-CNN model.In the first layer of the hierarchical model,the networks and the classifiers are trained on the whole training sets.While in the second layer,the networks and the classifiers are trained on each cluster.It can distinguish the confusing classes better and improve the classification accuracy of the whole dataset.In this thesis,experiments were carried out on three datasets: CUB-200-2011,FGVC-Aircraft-2013 b and Stanford-cars.The accuracy of classification was increased from 84.35%,83.56% and 89.45% of the B-CNN model to 84.67%,84.11 %,89.78%,which verifies the validity of the algorithm.In addition,we have also done some other experiments and get some conclusions.At last,we put forward some ideas for future work.