Fast and Accurate Single Image Super-Resolution via Information Distillation Network

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. Code is available at https://github.com/Zheng222/IDN-Caffe. Recently, due to the strength of deep convolutional neural network (CNN), many CNN-based SR methods try to train a deep network to gain better reconstruction performance. Kim et al. propose a 20-layer CNN model known as VDSR [12], which adopts residual learning and adaptive gradient clipping to ease training difficulty. To control the model parameters, the authors construct a deeply-recursive convolutional network (DRCN) [13] by adopting recursive layer. To mitigate training difficulty, Mao et al. propose a very deep residual encoder-decoder network (RED) [17], which consists of a series of convolutional and subsequent transposed convolution layers to learn end-to-end mappings from the LR images to the ground truths. Tai et al. propose a deep recursive residual network (DRRN) [22], which em-arXiv:1803.09454v1 [cs.CV]

Lightweight Image Super-Resolution with Information Multi-distillation Network

“…where H I M D N (·) is our IMDN. It is optimized with mean absolute error (MAE) loss followed most of previous works [2,11,18,36,38]. Given a training set…”

“…Very recently, Li et al [17] exploited feedback mechanism that enhancing low-level representation with high-level ones. For lightweight networks, Hui et al [11] developed the information distillation network for better exploiting hierarchical features by separation processing of the current feature maps. And Ahn [2] designed an architecture that implemented a cascading mechanism on a residual network to boost the performance.…”

Learned Multi-View Texture Super-Resolution

“…Section 4.3), which may be problematic if the brightness varies across input views. Table 1 presents a quantitative comparison of our method with two baselines (for an upsampling factor ×4): an reimplementation of [55] with the more robust L 1 -dataterm, and the original version of the single-image SR network [22]. It also shows respective performance of MVA, SIP and MVA+SIP network.…”

“…Quantitative comparison of different texture super-resolution techniques (upscaling factor ×4, same initial texture atlas, all computed on the Y-channel images). The top two rows are baselines : our reimplementation of [55] with primal-dual optimization scheme and L 1 -dataterm (first row), single-image super-resolution network [22] (second row). We evaluate the individual components of our proposed approach: MVA subnet trained on our data, which is very similar to [55], but estimates the local blur from the data (third row); SIP subnet alone trained only on DIV2K (fourth row); and our complete network MVA+SIP (last two rows).…”

“…Ablation study and comparison (upsampling factor ×4). We compare our result (fourth column) to: LR input view (second column), ground truth obtained from [55] with upsampling factor ×2 from the original images (third column); output of the MVA subnet (fifth column); output of the SIP subnet without MVA (sixth column); output texture of [58] from input views super-resolved with [22] (seventh column); and output texture directly from LR input views with [58] (eighth column).…”