Research For Fine-grained Image Classification Based On Deep Learning

With the rapid development of modern technology,most of the existing image classification technology is based on deep learning technology and has achieved very good results in the classification and recognition of fine-grained objects.Classification and recognition of finegrained objects are mainly based on the identification of different types of objects,such as people,cars and dogs on the road.Its focus is on the differences between different types of objects.However,there is less attention to the specific brand of vehicle and the specific breed of dog,which is the same kind of species,so it can not be better satisfied for Identify and categorize requirements at different levels.With the growing maturity of fine-grained image classification technology,fine-grained image classification and recognition has gradually become the one of the main research directions of scientists.The challenges of fine-grained image classification are mainly manifested in subtle inter-class differences and large intra-class differences.The existing research methods focus on using BCNN to extract deeper aggregate dimensional feature representation methods and extracting discriminative parts of the image for classification and recognition,The two methods have their own advantages and disadvantages in recognition accuracy and calculation efficiency.This paper studies these two kinds of methods in depth and puts forward its own improvement methods.Experiments are carried out on the fine-grained image representative dataset CUB-200-2011 bird dataset,Stanford Cars vehicle dataset and FGVC Aircraft dataset.,to prove the practicability of the method,a large-scale Chinese bird database is built and an application system for avian classification and recognition is developed.The main research work of this paper is as follows:(1)Propose a classification model based on low-dimensional hybrid bilinear network.Using bilinear model(B-CNN)to obtain high-dimensional aggregate features of fine-grained images for image classification and recognition has been proved to be an effective method.Commonly used fine-grained image classification methods do not connect the image location and features end-to-end,but B-CNN can combine them organically.However,traditional BCNN models usually use two identical types of convolution neural networks,such as VGG or Res Net.This paper proposes using different types of hybrid convolution neural networks.VGG networks focus more on object location and Res Net networks focus more on feature information extraction.Meanwhile,B-CNN model is time-consuming and memory-intensive.In order to overcome the shortcomings of B-CNN model,the principal component analysis(PCA)method is proposed in this paper.The key image information of bilinear features is retained for classification and recognition.Finally,the experiments show that the classification accuracy of hybrid bilinear network model is better than that of the same type of network model.At the same time,the PCA dimensionality reduction method can reduce the computational complexity and parameters while still maintaining a good classification effect.(2)Propose a method of classification based on image key regions.There are many local features of fine-grained images,some of which have a significant impact on the final classification results,and some of which have no impact on the final classification results.It is very important to select the key discriminant image regions.The traditional method is to extract key parts based on labeling designed by human.This method is time-consuming and labor-consuming,and the key parts of labeling are not necessarily accurate.Existing research mostly extracts key regions by designing constraint functions.In this paper,a method of reinforcement learning is proposed.Different image regions are selected iteratively,and the classification success rate is used as a reward function.Discriminative-Region-Net(DR-Net)is a network model for classification and recognition of key image regions.Experiments show that the model n achieves very good recognition results on standard data sets.(3)Design and implement Chinese bird classification and recognition application System.In order to verify the practicability of the method,this paper takes Chinese birds as an example,grabs nearly 1 million bird images from the Internet,eliminates dirty data through clustering analysis,and builds nearly 500,000 Chinese bird image data sets.According to the DR-Net deep learning model proposed in this paper,a Chinese bird recognition application system is designed and implemented based on the self-built Chinese bird data set.