Research On Retrieval Methods Of Multi-label Images Using Residual Networks

With the rapid development of Artificial Intelligence,Internet of Things and multimedia technologies,the number of images is growing explosively.How to efficiently and accurately retrieve the images required by users is a key issue in the field of multi-label image retrieval.In order to improve the accuracy and efficiency of multi-label image retrieval,the deep hashing methods are adopted for the mainstream solution.The deep hashing methods have achieved certain achievements in the field of multi-label image retrieval,however,there are still some prominent problems such as low accuracy in the field of multi-label image retrieval.In this dissertation,we focus on the relevant issues and present a multi-label image retrieval method based on residual network,mainly including the following three aspects of work:(1)In the existing deep network models,the extracted low-level features cannot effectively integrate the multi-level semantic information and the similarity ranking information of pairwise multi-label images into one hash coding learning model.Therefore,in this dissertation,we construct a deep hashing model based on improved residual network.The cosine distance of pairwise multi-label images label vector is adopted for the model to quantify existing multi-level similarity information in a multi-label image.Meanwhile,we utilize the residual network to improve the learning ability of the model.Extensive experiments on two popular multi-label datasets demonstrate that the improved model outperforms the reference best benchmark model regarding accuracy.The mean average precision is improved by 4.09% and 8.47% on two datasets,respectively.(2)In the existing deep hashing models,the extracted features vector usually contain global information of the multi-label images,including object information of the multi-label images and cluttered background information.Therefore,in this dissertation,we construct a deep hashing model for residual networks with attention mechanism.An attention mechanism is introduced in the model to identify the approximate location of the object(foreground)in a multi-label image.Simultaneously,in order to adapt to the model learning based on the attention mechanism,we modify the loss function based on the third chapter of the dissertation.Extensive experiments on two popular multi-label datasets demonstrate that the improved model outperforms the reference best benchmark model regarding accuracy.The mean average precision is improved by 5.29% and 9.83% on two datasets,respectively.(3)Due to the limited computing power and resources of single GPU,the training speed of deep residual network model is slow.After analyzing the advantages and disadvantages of data parallelism and model parallelism,we design a multi-GPU data parallel method based on Keras framework to improve training efficiency according to the improved deep residual network model,so as to solve the problem of the improved deep residual network model has long training time on the single machine.Through comparison experiments,it is verified that the multi-label image retrieval method based on residual network has higher accuracy.