Abstract
In recent years, deep convolutional neural networks (CNNs) have been widely used for image super-resolution (SR) to achieve a range of sophisticated performances. Despite the significant advancement made in CNNs, it is still difficult to apply CNNs to practical SR applications due to enormous computations of deep convolutions. In this paper, we propose two lightweight deep neural networks using depthwise separable convolution for the real-time image SR. Specifically, depthwise separable convolution divides the standard convolution into depthwise convolution and pointwise convolution to significantly reduce the number of model parameters and multiplication operations. Moreover, recursive learning is adopted to increase the depth and receptive field of the network in order to improve the SR quality without increasing the model parameters. Finally, we propose a novel technique called Super-Sampling (SS) to learn more abundant high-resolution information by over-sampling the output image followed by adaptive down-sampling. The proposed two models, named SSNet-M and SSNet, outperform the existing state-of-the-art real-time image SR networks, including SRCNN, FSRCNN, ESPCN, and VDSR, in terms of model complexity, and subjective and PSNR/SSIM evaluations on Set5, Set14, B100, Urban100, and Manga109.
Highlights
Image super-resolution (SR) is to convert an observed low resolution (LR) into a high resolution image (HR) by adding plausible high-frequency information to improve the visual quality according to perception of the human visual system
Our main contributions are as follows: - We introduce the first recursive depthwise separable convolution network for image super-resolution, in order to formulate the real-time models with extremely low complexity, which improves the image quality while maintaining the same level of model parameters. - We propose a novel technique called as super-sampling for generating the higher quality HR image for superresolution
Due to the wide applications of real-time image superresolution in various research areas, it is highly desirable to propose new algorithms for improving the existing realtime image super-resolution networks based on deep learning
Summary
Image super-resolution (SR) is to convert an observed low resolution (LR) into a high resolution image (HR) by adding plausible high-frequency information to improve the visual quality according to perception of the human visual system. Bicubic interpolation has been widely adopted as the traditional method for image SR by using cubic polynomials to model the image signals, in order to interpolate the missing HR pixels [2]. The abrupt signal changes in natural images can hardly be modeled by 3rd order polynomials, such that adaptive edge-directed interpolation methods were proposed to model the edge characteristics to better reconstruct the edges [2]–[4]. Due to the rapid development of computer hardware, the processing power of computer has reached a certain level that learning-based approaches were widely developed to explicitly learn the relationship of the LR and HR images using the handcrafted models, which often involve well-designed components, including feature extraction, non-linear mapping and reconstruction, using large external database or sparse representations, etc, as the source of information for offline supervised learning [5]–[7]
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
