Structure-preserving image smoothing with semantic cues
- PDF / 18,253,304 Bytes
- 11 Pages / 595.276 x 790.866 pts Page_size
- 59 Downloads / 189 Views
ORIGINAL ARTICLE
Structure-preserving image smoothing with semantic cues Linggang Chen1 · Gang Fu2
© Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract The purpose of image smoothing is to smooth out low-contrast textures while preserving meaningful structures. Although this problem has been studied for decades, it still leaves a lot of space to improve. Recently, learning-based edge detectors have superior performance to traditional manually-designed detectors. Based on the edge detection technique, we present a novel optimization-based image smoothing model combining semantic prior and perform L 0 gradient minimization recursively in our framework to refine the result. Our framework combines the advantage of the state-of-the-art edge detector and the ability of L 0 gradient minimization for structure-preserving image smoothing. Moreover, we employ a large number of real-world images and perform various experiments to evaluate our algorithm. Experimental results show that our algorithm outperforms state-of-the-art algorithms, especially in extracting subjectively-meaningful structures. Keywords Structure-preserving smoothing · Texture · Median filtering
1 Introduction looseness-1Structure-preserving image smoothing is a fundamental and important problem in computer vision and computer graphics. It aims at removing low-contrast textures/noises while preserving meaningful structures as much as possible. Moreover, this technique encourages many applications, e.g., edge enhancement as well as extraction, clip-art restoration, depth upsampling, object classification and many other image editing tasks [44]. looseness-1A wide variety of previous structure-preserving image smoothing techniques often depend on gradient-based measure to extract structures from textures, e.g., L 0 gradient measure [44], relative total variation measure [45] and relativity-of-Gaussian (RoG) [5]. These traditional methods sometimes fail to remove small-scale textures/details that have a relatively small effect on the semantically meaningful structures. However, we notice that, the L 0 minimization has been extended to mesh denoising [16] and point set denois-
B
Gang Fu [email protected] Linggang Chen [email protected]
1
Yunzhangfang Network Technology Co., Ltd., Nanjing 210000, China
2
School of Computer Science, Wuhan University, Wuhan 430072, China
ing [38], due to its the simplicity and the effectiveness of low-contrast texture removal. looseness-2For another thing, due to the development of deep learning, significant progress has been made in the edge detection community over the past few years. It has been shown recently that the results estimated by learning-based edge detectors are consistent with the human visual system [8,25,43]. Existing state-of-the-art edge detection method can powerfully extract the edges of objects with different sizes from textures/noises, which provides us with useful semantic cues. Intuitively speaking, an ideal edge detection result is thus a guidance for structure-preserving image
Data Loading...
