Learned Adaptive Kernels for High-Fidelity Image Downscaling

TL;DR

This paper presents ADK-Net, a deep learning framework that predicts spatially-varying, channel-specific kernels for high-fidelity image downscaling, outperforming existing methods in preserving image quality.

Contribution

Introduction of ADK-Net, a novel neural network that learns adaptive, pixel-wise, and channel-specific kernels for superior image downscaling fidelity.

Findings

Achieves state-of-the-art PSNR and SSIM on benchmark datasets.

Effectively models channel interdependencies for improved downscaling.

Outperforms traditional and existing learning-based methods.

Abstract

Image downscaling is a fundamental operation in image processing, crucial for adapting high-resolution content to various display and storage constraints. While classic methods often introduce blurring or aliasing, recent learning-based approaches offer improved adaptivity. However, achieving maximal fidelity against ground-truth low-resolution (LR) images, particularly by accounting for channel-specific characteristics, remains an open challenge. This paper introduces ADK-Net (Adaptive Downscaling Kernel Network), a novel deep convolutional neural network framework for high-fidelity supervised image downscaling. ADK-Net explicitly addresses channel interdependencies by learning to predict spatially-varying, adaptive resampling kernels independently for each pixel and uniquely for each color channel (RGB). The architecture employs a hierarchical design featuring a ResNet-based feature extractor and parallel channel-specific kernel generators, themselves composed of ResNet-based trunk and branch sub-modules, enabling fine-grained kernel prediction. Trained end-to-end using an L1 reconstruction loss against ground-truth LR data, ADK-Net effectively learns the target downscaling transformation. Extensive quantitative and qualitative experiments on standard benchmarks, including the RealSR dataset, demonstrate that ADK-Net establishes a new state-of-the-art in supervised image downscaling, yielding significant improvements in PSNR and SSIM metrics compared to existing learning-based and traditional methods.