SepLUT: Separable Image-Adaptive Lookup Tables for Real-Time Image Enhancement

AbstractImage-adaptive lookup tables (LUTs) have achieved great success in real-time image enhancement tasks due to their high efficiency for modeling color transforms. However, they embed the complete transform, including component-independent and component-correlated parts, into only a single type of LUTs, either 1D or 3D, coupled manner. This scheme raises dilemma improving model expressiveness two factors. On one hand, LUTs provide computational but lack critical capability components interaction. other, 3D present enhanced transform suffer from heavy memory footprint, training difficulty, limited cell utilization. Inspired by conventional divide-and-conquer practice signal processor, we SepLUT (separable image-adaptive table) tackle above limitations. Specifically, separate cascade sub-transforms instantiated as respectively. In this way, capabilities can facilitate each where LUT complements ability mix up components, redistributes input colors increase utilization thus enable use more lightweight LUT. Experiments demonstrate that proposed method presents performance on photo retouching benchmark datasets than current state-of-the-art achieves processing both GPUs CPUs.