A non-smooth non-local variational approach to saliency detection in real time
- PDF / 1,692,283 Bytes
- 12 Pages / 595.276 x 790.866 pts Page_size
- 1 Downloads / 170 Views
ORIGINAL RESEARCH PAPER
A non‑smooth non‑local variational approach to saliency detection in real time Eduardo Alcaín1,2 · Ana I. Muñoz1 · Emanuele Schiavi1 · Antonio S. Montemayor2 Received: 27 December 2019 / Accepted: 4 September 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract In this paper, we propose and solve numerically a general non-smooth, non-local variational model to tackle the saliency detection problem in natural images. In order to overcome the typical drawback of the non-local methods in image processing, which mainly is the inherent computational complexity of non-local calculus, as the non-local derivatives are computed w.r.t every point of the domain, we propose a different scenario. We present a novel convex energy minimization problem in the feature space, which is efficiently solved by means of a non-local primal-dual method. Several implementations and discussions are presented taking care of the computing platforms, CPU and GPU, achieving up to 33 fps and 62 fps respectively for 300×400 image resolution, making the method eligible for real time applications. Keywords Variational methods · Convex · Primal-dual · Non-local image processing · Saliency segmentation · GPU · Superpixels
1 Introduction Saliency object segmentation refers to an image processing system which aims to emulate the human visual attention system extracting the most relevant information of a scene. Saliency methods are usually applied as a pre-processing step in different areas in computer vision: adaptive compression of images [37], image retrieval [14] and content-aware resizing [4] and can be divided into bottom-up (pre-attentive data driven) and top-down (task dependent). In this work, we shall focus on bottom-up, image stimulus-driven models of attention. They are task-free and do not rely on learning, training or contextual information. They can also be * Antonio S. Montemayor [email protected] Eduardo Alcaín [email protected] Ana I. Muñoz [email protected] Emanuele Schiavi [email protected] 1
Department of Applied Mathematics, Universidad Rey Juan Carlos, C/Tulipán, S/N, 28933 Móstoles, Spain
Department of Computer Science and Statistics, Universidad Rey Juan Carlos, C/Tulipán, S/N, 28933 Móstoles, Spain
2
considered as fundamental building blocks for advanced, robust, hybrid bottom-up and top-down models. Several algorithms and methods have emerged in the classical image processing field for the saliency detection problem like [1, 10], machine learning approaches [17, 36], variational methods [24, 29] and combined [15, 18]. In Itti et al. [20], the saliency map is determined by using center-surround operations on colour, intensity and orientation features using a Difference of Gaussians (DoG) approach in a multiscale framework. In [1] the authors use the L∗a∗b colour space, subtracting the image mean colour to each of the components and producing the saliency map after a meanshift segmentation and a dynamic thresholding. As a segmentation
Data Loading...
