Deep Super-Resolution Hashing Network for Low-Resolution Image Retrieval
| Autor: | Qian Wang, Xiaobin Zhu, Naiguang Zhang, Zhuangzi Li, Feng Dai, Peng Li |
|---|---|
| Rok vydání: | 2019 |
| Předmět: |
Computer science
business.industry Deep learning Hash function ComputingMethodologies_IMAGEPROCESSINGANDCOMPUTERVISION Pattern recognition 02 engineering and technology Iterative reconstruction 010501 environmental sciences Resolution (logic) 01 natural sciences Superresolution Discriminative model 0202 electrical engineering electronic engineering information engineering 020201 artificial intelligence & image processing Code generation Artificial intelligence business Image retrieval 0105 earth and related environmental sciences |
| Zdroj: | Lecture Notes in Computer Science ISBN: 9783030341121 ICIG (3) |
| DOI: | 10.1007/978-3-030-34113-8_39 |
| Popis: | In image retrieval, deep learning based hashing approaches have achieved promising performance in recent years. However, they are usually trained on a specific resolution, so it will cause unpleasant retrieval results with low-resolution input. In this paper, we propose a novel end-to-end deep super-resolution hashing network (DSRHN) for low-resolution image retrieval. It aims to adopt super-resolution techniques to promote semantic information of low-resolution images, so that benefits hashing code generation in a more representative fashion. The proposed network consists of two major components, which are trained alternatively, named super-resolution network and hashing network. The super-resolution network is not only optimized by MSE loss for pixel-wise image reconstruction, but also optimized by perceptual loss extracted by the hashing network for semantic learning. As for the hashing network, we adopt hashing semantic loss to optimize it for accurate hash code generation, and utilize a discriminative loss to improve the discriminative ability for the super-resolved images and high-resolution images. Extensive experiments show that our method achieve state-of-the-art performance on low-resolution images retrieval. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|