Multi-label Image Classification And Its Application In Remote Sensing Image Recognition

In the multi-label image classification task,it is a key problem to improve the recognition accuracy of the model by mining label correlation.At present,the most advanced method is to use the label co-occurrence frequency of the target data set to establish a complete graph convolutional network.However,constructing such a global graph for the entire dataset may cause the frequency-bias and reduce the generalization ability of the model.Therefore,this paper proposes an attention-driven multi-label dynamic graph convolutional network ML-DGCN,which generates specific dynamic graphs for each image to enhance the robustness of learning features.In addition,we find that in the field of remote sensing image recognition,multi-label remote sensing image recognition is actually a multi-label learning problem.Therefore,this paper applies the proposed model to multi-label remote sensing image classification,and improves the model according to the common problem of long-tailed distributions in remote sensing data sets,and achieves good results.The following is a summary of the work content and innovation of the paper:(1)In order to fully mine label correlation and solve the problem of frequency-bias caused by label co-occurrence,this paper proposes dynamic graph convolutional network ML-DGCN.In the semantic category module of ML-DGCN,we locate the semantic regions of specific categories in the image based on class activation mapping,and then use the activation map to decompose the convolutional feature map into multiple content-aware category representations.In the classifier learning module,the input of the graph convolutional network is no longer the external word embedding vector,but takes the content-aware category representation as the input node of the graph convolutional network.The content-aware category representation passes through a global static graph that captures coarse label dependencies and a local dynamic graph that captures fine label dependencies.Joint graph convolutional network uses different graph nodes to propagate label semantic information and generate classifiers for images.(2)In the field of multi-label remote sensing image recognition,a sample will correspond to multiple object level labels at the same time.We believe that it is also a multi-label classification problem in essence.Therefore,this paper applies the proposed ML-DGCN model to multi-label remote sensing image recognition.The graph convolutional network is used to make up for the lack of semantic information correlation in multi-label remote sensing images and improve the classification ability of the model.In addition,for the common long-tailed distributions in remote sensing image data sets,this paper takes the measure of changing the loss function,and sets different weights for positive and negative samples,so as to improve the contribution of tail classes to the model.The results show that this method can effectively alleviate the long-tailed distributions of remote sensing data sets,and has a good performance in relevant evaluation indexes.Experiments show that the model in this paper can achieve good classification results compared with the existing methods,whether in the traditional multi-label image classification scene or in the field of multi-label remote sensing image recognition.