Research And Application Of Multi-label Classification Algorithm Based On Deep Learning

Multi-label classification is an emerging technology for processing multi-label classification of complex data,which has been widely used in text classification,semantic image annotation,bioinformatics analysis,and audio sentiment detection.The multi-label classification method based on deep learning can extract high-level abstract features,with strong adaptability and high accuracy.However,in the actual application of such methods,there are still problems such as network overfitting,insufficient extraction of important labels,and low correlation between labels.Therefore,this paper proposes two improved methods to further improve the accuracy of multi-label classification models based on deep learning,including the following work:Aiming at the network overfitting problem of multi-label classification algorithm based on deep learning,a hybrid multi-label classification model based on multi-attention mechanism: YMLC model,is proposed.The model applies the hybrid attention mechanism HPMI to the YOLOv5s network model,and the feature weights obtained in the channel dimension and spatial dimension are weighted and fused to obtain the final output weight through the multi-branch structure of HPMI,so that it can fully learn the weight distribution of channel domain and spatial domain,so as to not only pay attention to the position of important features,but also improve the feature expression ability of key regions.The model makes full use of the complementary features of channel attention mechanism and spatial attention mechanism in two different dimensions,promotes the information flow between key image information and network model,and improves the information utilization rate of the network training set,thereby alleviating the overfitting phenomenon of the network.The experimental results show that the recognition accuracy of the improved model in the COCO dataset is improved by 0.28 compared with the original network model.Aiming at the problems of insufficient extraction of important labels and low correlation between labels in the feature extraction process of traditional multi-label classification models,an LCAM model is constructed.The LCAM model mainly includes a feature extraction module and a classifier module.In the feature extraction module,this paper adds the Triplet attention mechanism to the VGG network model,which fully extracts the feature information in the channel dimension and spatial dimension with three branch structures,so as to solve the problem that the convolutional neural network ignores the different importance of accurate recognition of all labels in different image regions in the process of feature extraction of images.In the classifier module,a label-related classifier is built for each label through a graph convolutional network to mine the dependencies between labels.Specifically,the classifier of each label is constructed by using the feature vector as the node of the graph and the label correlation matrix as the edge of the graph,and the correlation between the labels is fully explored by updating the node information.The experimental results show that the LCAM model improves the accuracy of the current optimal algorithm Multi-Evidence by 0.018 in COCO dataset,and also has higher recognition accuracy on RAF-ML face expression data,and 0.2 higher accuracy than the DBM-CNN algorithm.