Patch-based Co-occurrence Filter with Fast Adaptive Kernel

Zhonggui Sun1,2, Tingting Liu1, Jie Li2, Ying Wang2, Xinbo Gao3

1Liaocheng University, 2Xidian University,  3Chongqing University of Posts and Telecommunications


Which kind of edges should be preserved in texture smoothing, that is a major concern for experts. More recently, co-occurrence filter (CoF) gives a refreshing answer: in contrast to ordinary edges, the prominent structure edges (i.e., boundaries) ought to be given a prior consideration. In texture smoothing, CoF employs co-occurrence matrix to separate different edges in an image to preserve structures. However, the collection of co-occurrence information in the filter is pixel-wise, which is fragile to noises. To overcome the drawback, we extend CoF to a novel version, named patch-based co-occurrence filter with fast adaptive kernel (FAKPCoF). Different from acquiring pixel-wise co-occurrence matrix, FAKPCoF collects the co-occurrence information based on patches, which is more robust to noises. And rather than spatial Gaussian in CoF, an adaptive kernel is employed to keep the sharpness of edges. Meanwhile, to break through the computational bottleneck induced by the adaptive operation, a fast variant of the adaptive kernel is designed. Thorough experimental results demonstrate that the proposed FAKPCoF reveals more powerful performance both in texture smoothing and structure-preserving.


Z. Sun, T. Liu, J. Li, Y. Wang, X.-B. Gao. "Patch-based Co-occurrence Filter with Fast Adaptive Kernel," in Signal Processing, 2021. DOI: 10.1016/j.sigpro.2021.108089 [pdf|code]







Visualization Results