End-to-End Feature Learning for Multi-label Image Retrieval

Hashing code similarity search is an effective method for large-scale image retrieval, since it provides fast search calculating and memory operating. In image retrieval, hash code based on deep convolutional networks can learn accurate image semantic information and compact hash code simultaneously. The existing methods generally extract the overall features of the image and generate corresponding hash codes. When the image contains multiple targets, this method may be suboptimal for describing the image, thus, limiting the accuracy of multi-label images. This paper focuses on the use of deep convolutional hash coding for multi-label image problems. In order to solve this problem, the authors propose a deep convolutional neural network structure suitable for multi-label image feature extraction. This network can also be used for end-to-end feature learning and further improve image retrieval accuracy. In this method, each target in the image is represented by a hash code and extensively evaluated across multiple data sets, making substantial improvement over the most advanced supervised and unsupervised hashing methods.