CURL: Neural Curve Layers for Global Image Enhancement

TPE: Lightweight Transformer Photo Enhancement Based on Curve Adjustment

“…Deep LPF [8] proposed spatially localized image enhancements through the use of automated filter parameters. CURL [9] designed a multicolor space loss function to globally adjust image attributes in a human-interpretable manner through joint training in three different color spaces. Zero-DCE [2] implemented unsupervised photo enhancement through a lightweight network combined with a set of nonreference loss functions.…”

“…The results of CURL [9] are generally brighter and have significant artifacts in the outline. DeepUPE [19] has good global image enhancements, but some of the enhancement results have unnatural colors.…”

“…This section describes how to accomplish the photo enhancement task by using a curve adjustment [9], [28], [29]. Two problems need to be solved; the first problem is obtaining the adjustment parameters through the encoder and the second is applying the adjustment parameters to the input image.…”

“…Under the MIT Adobe FiveK dataset [35], TPE was compared with Exposure [18], DIS-Rec [17], U-NET [37], DPE [33], CURL [9], DeepLPF [8], HDRNet [3] and DeepUPE [34] in terms of PSNR, SSIM and LPIPS. We tested the performance of different methods on low-resolution images (approximately 512×341) and the original resolution images (approximately 12 megapixels).…”

“…TABLE 3 shows FPS with different methods at different resolutions of MIT-Adobe FiveK [35]. Due to the large models of CURL [9] and DeepLPF [8], it is impossible to enhance the original resolution image on NVIDIA 3090. HDRNet [3] has the highest FPS at 512P resolution images, with the TPE being 37% lower.…”