Feature Channel Correlation Constraint For Image Classification

Today,artificial intelligence has become a field with a wide range of applications and rapid development.The deep learning method,as an excellent representation learning method,is one of the important branches in the field of artificial intelligence.In recent years,deep learning has been successfully applied in computer vision,natural language processing,search recommendation and other fields due to its high algorithm accuracy and excellent data-driven characteristics.Image classification technology,as a basic task in the field of computer vision,is of key significance for achieving high-precision and interpretable image understanding.Image classification technology also has a large number of practical application cases in daily life,industrial production,security monitoring and other fields,which saves manpower and improves production efficiency.With a widespread application of deep learning algorithms in image classification tasks,how to improve the performance of image classification deep learning algorithms has become an important and hot issue.The main research field of this paper is the convolutional neural network(CNN)model in image classification tasks.In CNN models,extracting discriminative image features is the key to improve the accuracy of image classification.Previous studies have made many improvements in model structure and loss function design,aiming to improve sample separately in feature space.However,these research ignore the prior relationship of categories in image classification tasks.The coupling relationship between model parameters and training categories is so strong,which leads to confusion of high-level features between different categories and increases the difficulty of optimization.This paper proposes a feature space constraint based on the correlation of category relations,which includes:(1)Decoupling the model parameters of the convolutional neural network from the category relationship through the channel attention mechanism,extracting the category relationship vector representation,thus the model can adaptively learn the relationship between the categories in the image classification task.(2)In order to further enhance the discriminative ability of image features,the loss is restricted by the category relationship,so that the vector distances obtained from samples of different categories are increased,and the vector distances obtained from samples of the same category are reduced.Finally the model will obtain image features with compactness in class and separation between classes.(3)In the calculation process of constraint loss,in order to make full use of the advantages of parallel calculation,the item-by-item distance calculation is optimized into a mathematically equivalent matrix multiplication operation,which improves the calculation efficiency.This paper has conducted a large number of experiments on five image classification data sets(MNIST,CIFAR-100,Tiny-ImageNet,Cars-196,Birds-200).The experimental results show that the method proposed in this paper has achieved better performance than currently widely used image classification methods.Further comparative experiments also show the effectiveness of the various modules proposed in this article.