Banding Detection via Adaptive Global Frequency Domain Analysis

Banding artifacts, characterized by ring-like bands or color steps in smooth gradients of images and videos post low-bitrate compression or quantization, degrade viewer experience. Existing detection approaches primarily leverage spatial domain features such as gradients, contrast, and entropy, neglecting inherent frequency domain characteristics of banding artifacts. Addressing these limitations, this paper introduces a novel detection method utilizing significant frequency domain features of banding artifacts. The proposed algorithm, based on the Haar wavelet transform for frequency domain decomposition and an Adaptive Global Frequency Domain Filter (AGFF), distinctly emphasizes the directional features of artifacts while capturing the interplay between low-frequency gradient information and high-frequency artifact details. Extensive evaluations on the Band2k dataset demonstrated superior performance with a 96.18% accuracy, and AUROC and AUPRC scores of 0.9944 and 0.9922 respectively, outperforming existing SOTA methods.