Fast image search with deep convolutional neural networks and efficient hashing codes

Approximate nearest neighbor search is a good method for large-scale image retrieval. We put forward an effective deep learning framework to generate binary hash codes for fast image retrieval after knowing the recent benefits of convolution neural networks (CNN). Our concept is that we can learn binary codes by using a hidden layer to present the latent concepts dominating the class labels when the data labels are usable. CNN also can be used to learn image representations. Other supervised methods require pair-wised inputs for binary code learning. However, our method can be used to learn hash codes and image representations in a point-by-point manner so it is suitable for large-scale datasets. Experimental results show that our method is better than several most advanced hashing algorithms on the CIFAR-10 and MNIST datasets. We will further demonstrate its scalability and efficiency on a largescale dataset with 1 million clothing images.