DualBLN: Dual Branch LUT-aware Network for Real-time Image Retouching

Xiang Zhang, Chengzhe Lu, Dawei Yan, Wei Dong, Qingsen Yan; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 2139-2155

Abstract


The 3D Lookup Table (3D LUT) is an efficient tool for image retouching tasks, which models non-linear 3D color transformations by sparsely sampling them into a discrete 3D lattice. We propose DualBLN (Dual Branch LUT-aware Network) which innovatively incorporates the data representing the color transformation of 3D LUT into the real-time retouching process, which forces the network to learn the adaptive weights and the multiple 3D LUTs with strong representation capability. The estimated adaptive weights not only consider the content of the raw input but also use the information of the learned 3D LUTs. Specifically, the network contains two branches for feature extraction from the input image and 3D LUTs, to regard the information of the image and the 3D LUTs, and generate the precise LUT fusion weights. In addition, to better integrate the features of the input image and the learned 3D LUTs, we employ bilinear pooling to solve the problem of feature information loss that occurs when fusing features from the dual branch network, avoiding the feature distortion caused by direct concatenation or summation. Extensive experiments on several datasets demonstrate the effectiveness of our work, which is also efficient in processing high-resolution images. Our approach is not limited to image retouching tasks, but can also be applied to other pairwise learning-based tasks with fairly good generality.

Related Material


[pdf] [supp] [code]
[bibtex]
@InProceedings{Zhang_2022_ACCV, author = {Zhang, Xiang and Lu, Chengzhe and Yan, Dawei and Dong, Wei and Yan, Qingsen}, title = {DualBLN: Dual Branch LUT-aware Network for Real-time Image Retouching}, booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)}, month = {December}, year = {2022}, pages = {2139-2155} }