A Saliency-based Method For Fine-grained Image Classification

Fine-grained image classification task has attracted more and more attention of researchers as a hot research direction in computer vision.It aims to divide a basic class of objects into more fine-grained subclasses.For fine-grained images,the differences between the images in some subclasses are very small,or the images in some of the same subclasses are very different.The parts that can be used to distinguish images are often on local area.In addition,variances in the pose,scale or rotation usually make the problem more difficult.Researchers have visualized the convolutional layer of convolutional neural network and find that the saliency map of the image can be obtained by adding up all the feature maps,and then the target can be located weakly supervised.Therefore,this method is applied to the object Localization in the fine-grained classification task.This article has conducted in-depth research on this basis.The main work is as follows:(1)In this paper,we have improved the method of locating targets through saliency maps.We propose a two-step fine-tuning target location method.We locate objects by the first fine-tuned classification network and parts by the second fine-tuned.(2)We modified the loss function.Adding the center loss function to the softmax loss function to form a joint loss function.(3)The method adopted in this paper is composed of three parts: Locating objects and its parts and generating new datasets through the two-step fine-tuning object localiza-tion algorithm;training the classification networks for each datasets and extracting the features;features training a classification model with all the features.Inspired by the bilinear model,we improve the model.We combine the second and third step into a joint network model and train it.In this paper,we test and compare with a large number of state-of-the-art methods on the CUB-200-2011,Stanford Dogs and FGVC-aircraft datasets.The results prove the effectiveness of the our algorithm.The comparison of the mixed loss function with the original loss function proves that the effectiveness of the center loss function.Our approach achieves comparable or better accuracy than other weakly supervised algorithms and achieves the best accuracy on the Stanford Pet Dog and FGVC-aircraft datasets.Our method is even better than some strong supervised methods.