Image Multi-label Classification Algorithm Based On Deep Learning

Image classification is one of the research hotspots in the field of computer vision,and the multi-label image classification is widely used in intelligent photo albums,image intelligent management and other fields.As the rapid development of information technology,there are countless pictures and video resources on the Internet,and the styles are different.How to manage these pictures with different categories and contents has become the focus of many scholars.Thanks to the development of deep learning,convolutional neural networks have been applied to image multi-label classification and achieved good results.However,there are multiple objects on the same image,what is more,the shape and position of the objects are random.Due to the complexity of the multi-label image,it is challenging to classify the multi-label images.The existing multi-label classification methods do not fully reflect the efficiency of the network structure,nor do they fully consider the category association between different objects,which makes the multi-label classification unable to achieve high accuracy.Based on this,this paper studies some image multi-label classification,the main work is as follows:(1)Based on the network structure of Refinedet,this paper proposes a densely connected refinement network for image multi-label classification,which is abbreviated DCRN.The dense connection module in DenseNet is added to the connection process of different scale feature maps of two Refinedet modules.So that each layer can be directly used by all the subsequent layers,which enhances the reuse of features of different scales throughout the network,and makes the model more concise.The kind of dense connection improves the gradient backpropagation,and makes the network easier to train.What is more,it improves the accuracy of the algorithm to a certain extent.(2)In this paper,we introduce the attention mechanism on the basis of DCRN,and propose the DCRN algorithm under the attention mechanism for multi-label classification.The attention mechanism can give different weights to the objects of different importance by learning the similarity between the objects.It assigns a large weight to the objects of concern and a smaller weight to the objects that are not important.In this way,the accuracy of recognition can be effectively enhanced.In this paper,the object relation module containing the attention mechanism is used instead of the non-maximum suppression(NMS)to remove duplicate detection boxes,it aims to avoid the problem that the NMS needs to set the parameters manually.By assigning different weights to the class scores and bounding box coordinates of different objects,the final class probability output and the bounding box regression are carried out.The experimental results show that this method makes the algorithm more accurate.(3)After researching and training the algorithm and obtaining higher accuracy results,this paper designs and implements an image multi-label classification system platform based on C/S architecture.Then we take the life photos which are taken by the author as the input of the system.And the image is labeled with several different categories.With the test,the task of image multi-label classification was completed very well.