Unsupervised Image Retrieval Based On Deep Features

In recent times,with the rapid development of electronic devices and social applications,all kinds of data information explosion growth.How to quickly and accurately find the images that users are interested in from the massive image data with much visual information has been the focus of image retrieval.In the field of e-commerce,famous e-commerce systems like Taobao allow users to retrieve similar items based on their uploaded images.In addition,well-known search engine companies such as Google and Baidu also have special image retrieval engines.With the outstanding performance of deep learning in computer vision,current image retrieval research pays more attention to the methods based on it,especially the deep convolutional neural network(CNN).As the features extracted from CNN are universal,depth features can be used for subsequent tasks.And now,many image retrieval methods based on depth feature of su-pervised learning have good performance,but they need manual annotation for the training data.And the retrieval results is heavily dependent on the quality of dataset,which leads to poor scalability.In view of the above problems,it can be seen that image retrieval in an unsupervised way has great potential.In this paper,we propose two methods based on depth features under unsupervised learning:1.Unsupervised image retrieval with mask-based prominent feature accu-mulation.Only the pre-trained CNN is used as a fixed feature extractor without fine-tuning CNN.As not all features of each channel and not all channels in the feature map are conducive to retrieval,in the paper,representative channels are selected by sorting of the prominent features accumulation with mask in each channel.This method especially shows improvement for images with complex backgrounds..2.Unsupervised image retrieval based on generative adversarial network(GAN).Here,the network has to be trained.limit the GAN input noise variable to binary and take the characteristics of each input image as a condition to learn the binary representation of each image,and generate a composite image similar to the original image at the same time.The generated image is then input to the discriminator for verification.To optimize the model,the self-attention layer and Wassertein distance are introduced to learn better hash code for retrieval.