Research Of Deep Learning For Fine-Grained Image Categorization

With the rapid development of mobile Internet and intelligent mobile,more and more images appear on the internet,including some generic images,such as architecture,and finegrained images,such as flowers,birds,etc.Comparing with the generic images,the existence of large within class differences,and large between-class similarities among the fine-grained images increases the difficulty of fine-grained image categorization;additionally,the growing number of images requires an algorithm that can quickly and easily train new model with little or without old dataset and the new model has the ability to identify the new dataset and old dataset,which is an incremental learning way.From the perspective of deep learning,this paper studies the fine-grained image categorization and incremental learning.The main contributions of this paper are summarized as follows:1.Based on deep learning and metric learning,a new fine-grained image categorization method,which is called improved center-loss method,is proposed.Following the centerloss method,our method increases a constraint term of the nearest other class center distance in the loss function,which has the advantage of increasing the distance between classes,so it not only has the advantage that center loss can minimize the distance in class,but also enlarge the distance between classes.Final experiment results in CUB200-2011 and Food-101 dataset indicate that our method can achieve the competitive performance for finegrained images classification.2.Based on deep learning and regularization,this paper proposes an incremental learning method which is called Reinforce Elastic Weight Consolidation(REWC).Based on the EWC method,by using L2 regularization,our method can have good classification performance in old dataset while training new model with a small number of samples.Final experiment results in Food-101 dataset indicate that our method has better classification performance for both old and new datasets.3.This paper collects a FOOD-SCUT image dataset.This dataset includes 70 kinds of food materials,which are commonly seen in our daily life,including vegetables and melons.In addition,in order to compare the experimental performance,this paper extracts different traditional features and use different classification methods,including convolutional neural networks to implement the food material recognition.