SepLUT: Separable Lookup Tables for Real-time Image Enhancement

Recently, Alibaba's audio‑video algorithm team and Shanghai Jiao Tong University released a joint paper “SepLUT: Separable Lookup Tables for Real‑time Image Enhancement”, accepted by ECCV 2022. The code and models are open‑source.

ECCV is a top computer‑vision conference; this year it received 5,803 submissions and accepted 1,650 papers (≈28% acceptance).

Background. Color enhancement is a fundamental ISP operation. Traditional ISP uses 1‑D and 3‑D lookup tables (LUTs) to approximate complex color transforms.

Motivation. Existing LUT‑based methods rely on manual tuning and single‑type LUTs, which cannot simultaneously model component‑independent and component‑correlated transformations efficiently.

Method. SepLUT decomposes a full‑color transform into a cascade of a 1‑D LUT (handling component‑independent adjustments such as brightness and contrast) and a 3‑D LUT (handling component‑correlated adjustments such as hue and saturation). A lightweight CNN processes a down‑sampled image to predict both LUTs via fully‑connected layers. The predicted LUTs are applied sequentially to the original image. The whole pipeline is trained end‑to‑end with an MSE reconstruction loss.

Quantization. Because LUT parameters live in the same color‑space as the output, they can be directly quantized or fixed‑pointed without noticeable quality loss, reducing model size and inference time.

Experiments. Ablation studies compare the learned 1‑D LUT with histogram equalization, showing superior adaptive color distribution. Visualizations confirm that the 1‑D LUT aligns image color distributions while the 3‑D LUT refines hue and saturation. Quantization experiments demonstrate negligible performance degradation. On the FiveK benchmark, SepLUT achieves state‑of‑the‑art objective metrics with low parameter count and real‑time speed.

Conclusion. By revisiting the classic divide‑and‑conquer ISP design, SepLUT provides an efficient, real‑time color‑enhancement solution and suggests that the separable LUT paradigm may benefit other vision tasks.