IBM3D: Integer BM3D for Efficient Image Denoising

Jingyu Yang, Xue Zhang, Huanjing Yue, Changrui Cai, Chunping Hou

Abstract: The block-matching collaborative filtering (BM3D) denoiser has been considered as a strong performer in image denoising, but it has high computational cost in block-matching and 3D transforms, which limits its practical applications, particularly in embedded video processing systems. In this paper, we propose an integer BM3D (IBM3D) that involves only integer operations. To integerize 3D transforms, the balance of approximation accuracy and denoising performance is carefully investigated for a wide range of noise levels. We propose an integer Wiener filter and investigate its performance over the original empirical Wiener filter with both analytical analysis and experimental verifications. The Kaiser window weighting is also integerized. The experiment results show that the proposed IBM3D provides comparable denoising performance to the original BM3D, and generates even better results for high noise levels. The proposed IBM3D requires less computation than the original BM3D, and can be deployed into embedded systems without or with limited floating-point computation resources, and ported to chips with smaller circuit areas and less power consumption.