From Fidelity to Visual Quality: A Semi-Supervised Approach for Low-Light Image Enhancement (Supplementary Material)

Wenhan Yang,Shiqi Wang,Yuming Fang,Yue Wang,Jiaying Liu

2020-01-01

Abstract:This supplementary material presents the ablation study of recursive band learning (the first stage of proposed deep recursive band network) and band recomposition (its second stage), the visualization results of the learned bands, and more visual comparisons. The compared methods include Bio-Inspired Multi-Exposure Fusion (BIMEF) [14], Brightness Preserving Dynamic Histogram Equalization (BPDHE) [5], Camera Response Model (CRM) [16], Differential value Histogram Equalization Contrast Enhacement (DHECE) [10], Dong [1], Exposure Fusion Framework (EFF) [15], Contrast Limited Adaptive Histogram Equalization (CLAHE) [17], Low-Light Image Enhancement via Illumination Map Estimation (LIME) [4], Multiple Fusion (MF) [2], Multiscale Retinex (MR) [7] Joint Enhancement and Denoising Method (JED) [11], Refined Retinex Model (RRM) [9], Simultaneous Reflectance and Illumination Estimation (SRIE) [3], Deep Retinex Decomposition (DRD) [13], Deep Underexposed Photo Enhancement (DeepUPE) [12], EnlightenGAN [6]. 1. Ablation Study for Recursive Band Learning We perform the ablation studies for Recursive Band Learning quantitatively. Five versions are compared. • Version 1 (RBL-v1) does not include the recursive processing (only once prediction) and multi-scale loss constraint, by replacing Eqn. (7) in the main submission as follows, LRect = −φ ( x̂Ts1 , x ) . (1) • Version 2 (RBL-v2) additionally includes the multi-scale loss constraint. • Version 3 (RBL-v3) includes the recursive processing but without the feature bypass connection and the recursive input, by replacing Eqn. (6) in the main submission as follows, [ Δf t s1 , Δf t s2 ,Δf t s3 ] = F t BLN F ( x̂t−1 s1 ) , f t si = Δf t si , i = 1, 2, 3, x̂ts3 = F t R s3 ( f t s3 ) , (2) x̂ts2 = F t R s2 ( f t s2 ) + FU ( x̂ts3 ) , x̂ts1 = F t R s1 ( f t s1 ) + FU ( x̂ts2 ) . • Version 4 (RBL-v4) includes the recursive processing and the recursive input by without the feature bypass connection, by replacing Eqn. (6) in the main submission as follows, [ Δf t s1 , Δf t s2 ,Δf t s3 ] = F t BLN F ( y, x̂t−1 s1 ) , f t si = Δf t si , i = 1, 2, 3, x̂ts3 = F t R s3 ( f t s3 ) , (3) x̂ts2 = F t R s2 ( f t s2 ) + FU ( x̂ts3 ) , x̂ts1 = F t R s1 ( f t s1 ) + FU ( x̂ts2 ) .