Progressive Perception-Oriented Network for Single Image Super-Resolution (1907.10399v2)

Abstract: Recently, it has been demonstrated that deep neural networks can significantly improve the performance of single image super-resolution (SISR). Numerous studies have concentrated on raising the quantitative quality of super-resolved (SR) images. However, these methods that target PSNR maximization usually produce blurred images at large upscaling factor. The introduction of generative adversarial networks (GANs) can mitigate this issue and show impressive results with synthetic high-frequency textures. Nevertheless, these GAN-based approaches always have a tendency to add fake textures and even artifacts to make the SR image of visually higher-resolution. In this paper, we propose a novel perceptual image super-resolution method that progressively generates visually high-quality results by constructing a stage-wise network. Specifically, the first phase concentrates on minimizing pixel-wise error, and the second stage utilizes the features extracted by the previous stage to pursue results with better structural retention. The final stage employs fine structure features distilled by the second phase to produce more realistic results. In this way, we can maintain the pixel, and structural level information in the perceptual image as much as possible. It is useful to note that the proposed method can build three types of images in a feed-forward process. Also, we explore a new generator that adopts multi-scale hierarchical features fusion. Extensive experiments on benchmark datasets show that our approach is superior to the state-of-the-art methods. Code is available at https://github.com/Zheng222/PPON.

• The paper presents a three-stage framework that first restores global content, then refines structural features, and finally synthesizes realistic textures in SISR.
• It utilizes a Hierarchical Feature Fusion Block with dilated convolutions and cascading Residual-In-Residual Fusion Blocks for effective multi-scale feature extraction.
• The proposed PPON outperforms conventional methods by achieving state-of-the-art PSNR and enhanced perceptual fidelity on benchmark datasets like Set5 and Urban100.

Progressive Perception-Oriented Network for Single Image Super-Resolution

The paper presents a novel approach to Single Image Super-Resolution (SISR) addressing the inherent challenges posed by traditional methods that solely focus on peak signal-to-noise ratio (PSNR) maximization and often result in visually unsatisfactory images at higher upscaling factors. The authors introduce the Progressive Perception-Oriented Network (PPON), which uniquely synthesizes high-quality images by progressively integrating content, structural, and perceptual details.

The principal innovation of this work lies in its three-stage network architecture. Initially, the network concentrates on minimizing pixel-level errors to restore global content. Subsequently, it utilizes these features to enhance structural retention in images. Finally, it employs fine-structure features for synthesizing realistic textures without introducing perceptual artifacts commonly associated with GAN-based methods. This stage-wise strategy facilitates the generation of super-resolved images that not only evaluate well on traditional metrics such as PSNR but also demonstrate superior perceptual quality.

The methodology incorporates a Hierarchical Feature Fusion Block (HFFB) leveraging dilated convolutions to capitalize on multi-scale data, thereby enhancing feature representation. Through a deep neural network with cascading Residual-In-Residual Fusion Blocks (RRFBs), the proposed model achieves state-of-the-art performance in benchmark datasets in terms of PSNR.

Quantitative results underscore the effectiveness of the proposed approach. The PPON outperforms contemporary SISR methods in visual fidelity, achieving superior numerical performance on established benchmark datasets such as Set5, Set14, BSD100, and Urban100. Importantly, it also demonstrates enhanced computational efficiency through a reduced parameter set and a fast training process utilizing progressive training phases.

In conclusion, the PPON demonstrates a synergistic integration of structural robustness and perceptual sophistication in image super-resolution tasks. It establishes an effective model for generating high-resolution images, advancing both practical applications and theoretical discourse. Future research may explore extending this progressive framework to related areas in computer vision, including video super-resolution and diverse image restoration tasks. The implications are broad, potentially enhancing applications across computational photography, medical imaging, and deep learning-based visual enhancement fields.