Fractional-Grey Wolf optimizer-based kernel weighted regression model for multi-view face video super resolution
- PDF / 3,520,670 Bytes
- 19 Pages / 595.276 x 790.866 pts Page_size
- 94 Downloads / 208 Views
ORIGINAL ARTICLE
Fractional-Grey Wolf optimizer-based kernel weighted regression model for multi-view face video super resolution Amar B. Deshmukh1 · N. Usha Rani1 Received: 27 October 2016 / Accepted: 5 December 2017 © Springer-Verlag GmbH Germany, part of Springer Nature 2017
Abstract Due to the advancement of the intelligent surveillance system in recent days, security and protection cameras are installed even in small shops, but the qualities of the image captured by surveillance cameras are low. The technique used for reconstruction of the high-resolution images from observed low-resolution image is called as super-resolution techniques. In order to alleviate the resolution problem and to provide desired information, fractional-Grey Wolf optimizer-based kernel weighted regression model is developed in this paper for multi-view face video super-resolution. Here, a new optimal kernel weight matrix for the interpolation of the super-resolution image is generated using the proposed FGWO algorithm, which is newly developed by integrating the GWO with fractional calculus. The experimentation of the proposed system is carried over UCSD face video databases, and the performance results are analyzed using SDME, PSNR, and SSIM with various existing techniques. The experimental results demonstrated that the proposed method improved the performance of super-resolution by achieving the maximum PSNR, SSIM and SDME value of 49.5909, 0.99 and 87.51 dB. Keywords Super-resolution · Optimal kernel · Fractional calculus · Fractional Grey Wolf optimizer (FGWO) · Second derivative-like measure of enhancement (SDME)
1 Introduction A significant way to reduce the complicated factors in surveillance videos is the video super-resolution [1]. The video-based super-resolution is used to generate high-resolution frames by complementing the information of pixels [2]. Video super-resolution [2–11] can be classified into three categories: (1) super-resolving each video frame via image super-resolution techniques; (2) reconstruction based method; (3) learning based or example-based methods. The latest and most effective single image super-resolution techniques [12–14] are used to super-resolving each frame in the first category. Reconstruction-based methods are used to construct the high-resolution frames using motion compensation technique and extract the information from multiple views of the same object along frames. In the reconstructionbased method, registration is an important step to reconstruct
* Amar B. Deshmukh [email protected] 1
Vignan University, Guntur, Andhra Pradesh 522213, India
the high-resolution frames. This method is suffered from the computational complexity. When the motions are complex, the reconstruction-based methods become invalid. Learning-based or example-based methods are used to generate the high-resolution frames using codebooks derived from keyframes in the mixed-resolution video. This method consists of two stages: the first one is learning and the second one is recovery. In learning stage, th
Data Loading...