Research On Semi-Supervised Deep Learning Methods For Finr-Grained Image Classification

Fine-grained image classification refers to the recognition of fine-grained subcategories of a coarse-grained base class.The application scenarios of fine-grained image classification include vehicle fine recognition,aircraft fine recognition,flower or plant fine recognition,vertical search,bird or dog fine recognition in animal protection,etc.Deep Convolutional neural network has achieved the best performance in many computer vision applications such as fine granularity image classification.There are two main reasons for deep Convolutional neural network to achieve these great achievements:one is to increase the complexity of the network model,but this will greatly increase the amount of computation,which will bring huge computational burden to computing equipment.At the same time,too large neural networks are not only difficult to train,but also lead to over fitting;The second is the emergence of large-scale annotated training sample sets,but in practical applications,creating a large-scale,high-quality,annotated training sample set is very time-consuming and expensive,and often requires relevant professional knowledge,especially for fine-grained images.However,through web crawlers or search engines,a large number of unlabeled samples can be quickly obtained from the internet.Therefore,studying semi supervised deep learning methods for fine-grained image classification has important practical significance and broad application prospects.The research work of this article mainly includes the following aspects:(1)To address the issue of increasing the complexity of the network model,which increases computational complexity and is prone to overfitting,an objective function based on the feature mean matrix is proposed to optimize the network model.The objective function based on the feature mean matrix can significantly improve network performance without increasing the complexity of the network model.Specifically,based on the image feature vectors extracted from the network model,a feature mean matrix is constructed,treating the column vectors in the feature mean matrix as class center vectors of different classes.Then,based on the feature mean matrix,a deep discriminative feature learning regularization term is designed.The function of this regularization term is to minimize the distance between the image feature vector and its corresponding class center vector,making the distance between different class center vectors as large as possible,thus making the feature vectors learned by the network have good intra class compactness and inter class separability.(2)On the basis of the feature mean matrix,a deep discriminative feature learning method for fine-grained images is proposed based on the class label hierarchy of fine-grained image datasets.Given a deep Convolutional neural network and a fine-grained image classification dataset with a hierarchical structure of class labels,firstly,in order to better model the hierarchical structure of class labels,it is proposed to replace the classification output layer of the original network with several new full connection layers group.Each full connection layers group corresponds to a class label level,and the proposed multistage Softmax Loss function is applied to these new layers for model training.Secondly,in order to fully utilize the hierarchical structure of class labels and the similarity relationship information between fine-grained class labels to supervise the network training process,it is proposed to construct deep discriminative feature learning regularization terms using class label hierarchical structure and multi-level feature mean matrix.(3)To solve the problem of high cost in constructing large-scale labeled data integration,a Agglomeration Degree transductive Semi-Supervised Learning method is proposed.The goal of semi-supervised deep learning is to use a small number of labeled and a large number of unlabeled samples to train a high-precision network model,so that it can approximate or even equal to the model accuracy that Supervised learning can achieve,thus greatly reducing the cost of training data labeling.The idea of transductive Semi-Supervised Learning is to regard the class labels of unlabeled training samples as variables that need to be learned and optimized.They are updated and optimized iteratively with model parameters during the training process until the model converges.(4)Combining the above objective function based on the feature mean matrix,the depth discrimination feature learning method for fine-grained images,and the Agglomeration Degree Transductive Semi-Supervised Learning method,a transductive Semi-Supervised depth learning method framework for fine-grained image classification is proposed.The principle of Transductive Semi-Supervised Learning is extended and applied to deep network training for fine-grained image classification tasks,Comprehensive experimental evaluation fully verifies the effectiveness of the proposed method in the task of semi supervised fine-grained image classification.