Deep Residual Net Based Compact Feature Representation for Image Retrieval

Deep learning technology has been introduced into many multimedia processing tasks, including multimedia retrieval. In this paper, we propose a deep residual net (ResNet) based compact feature representation improve the content-based image retrieval (CBIR) performance. The proposed method integrates ResNet and hashing networks to convert the raw images into binary codes. The binary codes of images in query set and that of the database are compared using Hamming distance for retrieval. Comprehensive experiments are executed on three public databases. The results show that the proposed method outperforms state-of-the-art methods. Furthermore, the impact of the deep convolutional network (DCNN)'s depth on the performance is investigated.