AttentionLUT: Attention Fusion-Based Canonical Polyadic LUT for Real-Time Image Enhancement

Recently, many algorithms have employed image-adaptive lookup tables (LUTs) to achieve real-time image enhancement. Nonetheless, a prevailing trend among existing methods has been the employment of linear combinations of basic LUTs to formulate image-adaptive LUTs, which limits the generalization ability of these methods. To address this limitation, we propose a novel framework named AttentionLut for real-time image enhancement, which utilizes the attention mechanism to generate image-adaptive LUTs. Our proposed framework consists of three lightweight modules. We begin by employing the global image context feature module to extract image-adaptive features. Subsequently, the attention fusion module integrates the image feature with the priori attention feature obtained during training to generate image-adaptive canonical polyadic tensors. Finally, the canonical polyadic reconstruction module is deployed to reconstruct image-adaptive residual 3DLUT, which is subsequently utilized for enhancing input images. Experiments on the benchmark MIT-Adobe FiveK dataset demonstrate that the proposed method achieves better enhancement performance quantitatively and qualitatively than the state-of-the-art methods.

1. INTRODUCTION

We very likely capture low-quality photos with advanced cameras or cell phones, which is because the quality of final photos is affected by external factors such as ambient light and temperature. To procure visually pleasing images, image enhancement [1], [2], [3], [4], [5], [6], [7] technology becomes indispensable for the refinement of these low-quality photographs. The realm of image enhancement has witnessed remarkable advancements through deep learning-based methodologies, consistently achieving state-of-the-art results. HDRNet [8] uses a low-resolution vision of the input image to predict a set of affine transformations in bilateral space and applies the upsampled vision of these transformations to enhance the input image. CSRNet [9] employs a conditional network to extract global features and utilizes 1 × 1 convolutions to enact pixel-independent transformations, which simulate global brightness adjustments and other enhancement operations. Zeng et al. [10] introduced a framework that predicts an image-adaptive 3DLUT, representing pixel-independent enhancement transformations, for real-time enhancing input images. Subsequent refinements in this domain include SA-3DLUT [11] and AdaInt [12], which respectively incorporate spatial information and image-adaptive sampling to improve the image-adaptive 3DLUT.