DeepLPF: Deep Local Parametric Filters for Image Enhancement

Abstract

Digital artists often improve the aesthetic quality of digital photographs through manual retouching. Beyond global adjustments, professional image editing programs provide local adjustment tools operating on specific parts of an image. Options include parametric (graduated, radial filters) and unconstrained brush tools. These highly expressive tools enable a diverse set of local image enhancements. However, their use can be time consuming, and requires artistic capability. State-of-the-art automated image enhancement approaches typically focus on learning pixellevel or global enhancements. The former can be noisy and lack interpretability, while the latter can fail to capture finegrained adjustments. In this paper, we introduce a novel approach to automatically enhance images using learned spatially local filters of three different types (Elliptical Filter, Graduated Filter, Polynomial Filter). We introduce a deep neural network, dubbed Deep Local Parametric Filters (DeepLPF), which regresses the parameters of these spatially localized filters that are then automatically applied to enhance the image. DeepLPF provides a natural form of model regularization and enables interpretable, intuitive adjustments that lead to visually pleasing results. We report on multiple benchmarks and show that DeepLPF produces state-of-the-art performance on two variants of the MIT-Adobe 5k [3] dataset, often using a fraction of the parameters required for competing methods.

Diverse and Adjustable Versatile Image Enhancer

“…By using a reinforcement learning, Park et al [30] proposed a global image modification model. To account for local adjustments, Moran et al [13] trained several local parametric functions. Similarly, Kim et al [12] proposed a two-stage approach that includes a channel-wise intensity transformation and local refinement network.…”

“…Finally, the total loss function utilized for the training process is given by: [11], UPE [14], and DLPF [13]) and five ER images. where γ rec , γ latent , γ QA , γ KL , γ PA , and γ adv are the hyperparameters to weight each term.…”

“…In the experiment, a training/testing split follows UPE [14], and nine recent state-of-the-art image enhancement methods were benchmarked: U-Net based model [47], HDRNet [23], White-Box [22], DPE [11], Distort-and-Recover [30], UPE [14], DLPF [13], Kim et al [12], and Zeng et al [31]. We employed three well-utilized metrics for quantitative comparison, namely: SSIM [50], LPIPS [54], and PSNR, between the enhanced images and GTs.…”

“…As can be seen in the table, all ER images seem to have similar preference scores. This implies that it is necessary to take into account various styles of enhancement to reflect the real-world user preference rather than depending on only one style (as done in most studies) [11], [13].…”

“…Deep-learning techniques [10] have been successfully applied in photograph enhancers that allow users to enhance images to expert-level images automatically [11], [12], [13], [14]. The MIT-Adobe-5K dataset [1] having 5K natural images and five corresponding expert-retouched (ER) images have been widely used for training them.…”