Research On Product Quantization Image Retrieval Based On Generative Adversarial Networ

Today,with the exponential growth of image data with people’s information interaction on the network,how to quickly and accurately retrieve the images is an urgent problem to be solved.At present,the image retrieval model based on deep learning can extract the semantic features of images.However,in the process of training the deep learning model,the training is often insufficient because of limited training data and data with complex backgrounds,so it is difficult for the neural network to learn the correct mapping relationship.After obtaining the semantic features of the image,the product quantization algorithm in the approximate nearest neighbor algorithm can reduce the dimension of the high-dimensional features,shorten the calculation time,and exclude some data of different categories in the calculation process to increase the retrieval efficiency.However,the k-means algorithm used in the current product quantization algorithm is locally optimal because of the random selection of its initial clustering center.To solve these problems,the main work of this paper is as follows:(1)In order to solve the problem of insufficient training caused by limited training data in the unsupervised image retrieval model based on deep learning,and the problem that it is difficult for complex images to learn the correct mapping relationship in deep learning,an unsupervised image retrieval method based on Optimized Generation Adversarial Network and convolution attention module is proposed.Firstly,the training data is expanded by using the generated countermeasure network to reduce the possibility of insufficient training.Then,the convolution attention module is added to the network of the model to enhance the feature extraction ability of the network.Finally,the accuracy of this part of the model in the image retrieval task is verified by the image retrieval task experiment in NUSWIDE and FLICKR25 K data sets,and the optimization of the above problems by the model proposed in this part is verified by the ablation experiment.(2)Aiming at the problem of local optimization of K-means algorithm used in product quantization algorithm,an unsupervised image retrieval based on random optimal product quantization is proposed.By using the random optimization product quantization algorithm for image features,and taking the average value through multiple experiments,the measurement standard of the quantization average level is obtained.The cluster center higher than the average level is selected for use,which relatively alleviates the instability brought by the local optimal band of K-means algorithm,and greatly increases the retrieval efficiency.Finally,the time efficiency of this part of the model in the image retrieval task is verified by the image retrieval task experiment in NUSWIDE and FLICKR25 K data sets.