Research On Image Classification Methods Based On Deep Learning

Image classification is an important and challenging research topic in the field of computer vision.Its technology aims to solve the problem of image search and image recognition.Given a specific image,image classification technology judges image category by identifying the information and content it contains.On the dataset with the image category label,it can find the correct category label for an input image to achieve the purpose of classification.With the continuous development of Internet technology,data has proliferated,and the traditional classification methods have been unable to meet massive data classification demands.The emergence of deep learning and the development of GPU hardware enable image classification technology to classify in an autonomous,intelligent direction.Deep learning training process integrates feature extraction,which replaces the process of manually extracting features,and supports computation and performance evaluation under large datasets.However,Images not only have the characteristics of large amount of information and rich expression,but also have a significant increase in the category and detail,which brings new challenge to image classification technology.Moreover,images can be classified into inter-class images and intra-class images.Low similarity and difference between inter-class has become another challenge in image classification.In addition,in the terms of datasets with image category labels,the deep network can classify images well.But when the images in the dataset are not labeled or the image categories do not exist in the dataset,the result are not very well and the network model does not have a good generalization ability,which is also an urgent problem to be solved in the field of image classification.In order to resolve these above problems,this paper starts from three aspects:scene image classification,image fine-grained classification and image classification technology generalization.The main contents are as follows:(1)Conducted comprehensive research and analysis on the current mainstream image classification algorithms.At first,this paper makes summary of related research development and technology and then deeply elaborated and analyzed from feature-based image classification methods,semantic-based classification methods and learning-based learning.Finally,the trend of image classification algorithm is predicted and forecasted.(2)Aiming at the problem of label ambiguity and computational cost in scene image classification,a Joint-CNN method based on Inception CNN(I-CNN)and Object Detection CNN(OD-CNN)is proposed.It uses OD-CNN to calculate the probability of each object in the image and generates a corresponding vector.At the same time,BN-Inception model was modified by I-CNN to classify images combined with OD-CNN.(3)Aiming at the problem of fine-grained image classification,a multi-region discrimination feature extraction based fine-grained classification method is proposed.The method extracts the features of key regions in the image by using multi-region discrimination feature extraction method and inputs the feature into the iterative network,and finally outputs the recognition result.(4)Aiming at the problem of low generalization ability of image classification model,a Zero-Shot Learning(ZSL)based classification generalization model is proposed.The method is based on iterative network to extract multi-scale features of the image and uses Semantically Consistent Regularization Part(SCoRe-Part)network structure to complete the projection of the visual semantic space,which effectively solves the problem of domain drift and improves the generalization ability of the classification model.The research in this paper improves the effect and practicability of image classification,which makes the technology have better development and application.Through a large number of research and analysis of its methods,this paper makes some contributions from the scene image classification,image fine-grained classification and classification model generalization.Experiments have been conducted on the relevant datasets and achieve state-of-the-art results.