Toward Robust and Efficient Low-Light Image Enhancement: Progressive Attentive Retinex Architecture Search

• DOI: 10.26599/tst.2022.9010017
• Publication Date: 2022-12-13T19:41:56Z
• Authors: Xiaoke Shang; Nan An; Shaomin Zhang; Nai Ding
• Keywords: Color Constancy; Macro

In recent years, learning-based low-light image enhancement methods have shown excellent performance, but the heuristic design adopted by most requires high engineering skills for developers, causing expensive inference costs that are unfriendly to hardware platform. To handle this issue, we propose automatically discover an efficient architecture, called progressive attentive Retinex network (PAR-Net). We define a new framework introducing attention mechanism strengthen structural representation. A multi-level search space containing micro-level on operation and macro-level cell is established realize meticulous construction. endow searched architecture with hardware-aware property, develop latency-constrained strategy successfully improves model capability explicitly expressing intrinsic relationship between different models defined in framework. Extensive quantitative qualitative experimental results fully justify superiority of our proposed approach against other state-of-the-art methods. series analytical evaluations performed illustrate validity algorithm.