[Paper Review] Fast and Accurate Single Image Super-Resolution via Information Distillation Network

Hui Zheng, Wang Xiu+1 more|arXiv (Cornell University)|Mar 26, 2018

Advanced Image Processing Techniques23 references59 citations

TL;DR

The paper introduces IDN, a lightweight CNN with information distillation blocks for fast and accurate single image super-resolution, achieving competitive PSNR/SSIM/IFC while significantly reducing computation time.

ABSTRACT

Recently, deep convolutional neural networks (CNNs) have been demonstrated remarkable progress on single image super-resolution. However, as the depth and width of the networks increase, CNN-based super-resolution methods have been faced with the challenges of computational complexity and memory consumption in practice. In order to solve the above questions, we propose a deep but compact convolutional network to directly reconstruct the high resolution image from the original low resolution image. In general, the proposed model consists of three parts, which are feature extraction block, stacked information distillation blocks and reconstruction block respectively. By combining an enhancement unit with a compression unit into a distillation block, the local long and short-path features can be effectively extracted. Specifically, the proposed enhancement unit mixes together two different types of features and the compression unit distills more useful information for the sequential blocks. In addition, the proposed network has the advantage of fast execution due to the comparatively few numbers of filters per layer and the use of group convolution. Experimental results demonstrate that the proposed method is superior to the state-of-the-art methods, especially in terms of time performance.

Motivation & Objective

• Motivate reducing computational cost and memory usage in SISR while preserving high reconstruction quality.
• Design a compact network (FBlock, DBlocks, RBlock) that progressively distills residual information.
• Introduce enhancement and compression units to extract and distill features efficiently.
• Train with MAE loss and fine-tune with MSE loss to improve reconstruction accuracy.

Proposed method

• Feature extraction block with two 3x3 convolutions to obtain initial features from LR input.
• Stacked information distillation blocks (DBlocks) each containing an enhancement unit (two shallow 3-layer paths with short/long-path feature handling) and a compression unit (1x1 conv) to distill information.
• Information distillation block mixes local short-path and long-path features to enrich representations, followed by compression to reduce redundancy.
• Global residual learning by adding the upsampled LR input to the reconstructed HR residual via a reconstruction block based on transposed convolution.
• Training uses MAE loss first and then fine-tunes with MSE loss for improved accuracy.

Experimental results

Research questions

• RQ1 Can IDN achieve real-time or near-real-time inference on standard datasets while maintaining high SR quality?
• RQ2 Do enhancement and compression units effectively distill and preserve useful information for progressive HR reconstruction?
• RQ3 How does IDN compare to state-of-the-art SR methods in PSNR, SSIM, IFC, and runtime across common benchmarks?

Key findings

• IDN achieves competitive PSNR/SSIM/IFC compared with state-of-the-art methods across Set5, Set14, BSD100, and Urban100, often ranking highly on PSNR and SSIM.
• IDN demonstrates superior inference speed, being substantially faster than methods like MemNet and DRRN (e.g., about 500x faster than MemNet on Urban100 with 2x upscaling).
• On Set5 x2, IDN achieves 37.83 dB PSNR and 0.9600 SSIM; x3 yields 34.11 dB PSNR and 0.9253 SSIM; x4 yields 31.82 dB PSNR and 0.8903 SSIM.
• On Set14 x2, IDN achieves 33.30 dB PSNR and 0.9148 SSIM; x3 29.99 dB PSNR and 0.8354 SSIM; x4 28.25 dB PSNR and 0.7730 SSIM.
• On BSD100 and Urban100, IDN generally outperforms several baselines in PSNR/SSIM, with IFC results indicating strong perceptual quality gains (as shown in Table 3).
• A compact 31-layer network with grouped convolutions and a global residual structure delivers real-time speed while maintaining competitive reconstruction accuracy.