Image smoothing combining edge-consistency with region-piecewise flatting

Abstract

Image smoothing constitutes a fundamental task within the realm of computer vision. Effectively preserving the structural integrity of edges during the image smoothing process represents an arduous challenge. Current methodologies persistently encounter issues in adequately exploiting and harnessing the intrinsic value of edge information within images, while simultaneously suffering from the limitation of employing overly simplistic constraints for achieving desirable smoothing outcomes. To solve these problems, we propose a new Edge Consistency and Region Piecewise Flatting Network (ECRPF-Net). To circumvent the impact of the convolution process on edge-preserving properties, our network adopts the approach of superimposing significant edge information onto the feature layer, while also incorporating a weak structure reinforcement module. To efficiently preserve the edge structure, we incorporate an edge consistency module (ECM), which utilizes the edge-response consistency relationship between the input and output images to ensure the preservation of edges. To enhance the quality of image smoothing and mitigate the occurrence of edge artifacts, we introduce the Region Piecewise Flatting Module (RPFM). This module partitions the image into different regions based on the similarities and differences in edge response features, and applies a flexible sub-region smoothing approach to constrain the final output, ensuring superior smoothing results. Experimental results demonstrate the excellent performance of the ECRPF-Net in preserving the prominent edge structure of the image, achieving more visual-appealing smoothing outcomes, and surpassing the majority of existing methods on public datasets.