Abstract: Deep convolutional neural network (CNN) makes great breakthroughs in computer vision. Recently, many works have demonstrated that the performance of the CNN depends on the stacked convolutional layers. It is obvious that the features of the fully connected layers lose the topological structures of images, and the convolutional layer features contain a large amount of redundant information that interferes with the performance of model. Thus, we propose an effective supervised deep Hashing method, enhancing convolutional deep Hashing (ECDH), which learns the binary codes from the strengthened convolutional layer. Specifically, an enhanced convolutional Hash layer is constructed between the top convolutional layer and the output layer, enhancing the local features of the convolutional layer outputs while learning the binary codes by optimizing an objective function. The proposed method works well for existing deep learning models such as Alex neural network (AlexNet), visual geometry group neural network (VGGNet), residual neural network (ResNet), and is easier to be trained. Compared with state-of-the-art methods, extensive experiments show that the proposed method achieves better retrieval performance.

Key words: CNN, Hashing, image retrieval, Hamming distance