Paper Title:Deep Learning for Single Image Super-Resolution:A Brief Review
I.Sections
section2:相关背景概念
section3:有效神经网络结构forSISR
section4:SR不同的用处
II.Backgrounds
1.
2.Interpolation-based SISR methods, such as bicubic interpolation
[10] and Lanczos resampling [11]
3.Learning-based SISR methods：
3.1.The Markov random field (MRF) [16] approach was first adopted by Freeman et al. to exploit the abundant real-world images to synthesize visually pleasing image textures.
3.2.Neighbor embedding methods [17] proposed by Chang et al. took advantage of similar local geometry between LR and HR to restore HR image patches.
3.3.Inspired by the sparse signal recovery theory [18], researchers applied sparse coding methods [19], [20], [21], [22], [23], [24] to SISR problems.
3.4.Lately, random forest [25] has also been used to achieve improvement in the reconstruction performance.
3.5.Meanwhile, many works combined the merits of reconstruction-based methods with the learning-based approaches to further reduce artifacts introduced by external training examples [26], [27], [28], [29].
4.deep learning:
artificial neural network (ANN)[32]->convolutional neural network (CNN) [35]->recurrent neural network (RNN) [36]->pretraining the DNN with restricted Boltzmann machine(RBM)->deep Boltzmann machine (DBM) [42], variational autoencoder (VAE) [43], [44] and generative adversarial nets (GAN) [45]
III.DEEP ARCHITECTURES FOR SISR
discuss the efficient architectures proposed for SISR in recent years.
A.Benchmark of Deep Architecture for SISR
- SRCNN: three-layer CNNCNN, where the filter
sizes of each layer are 64 * 1 * 9 * 9, 32 * 64 * 5 * 5 and 1 * 32 * 5 * 5.
loss function:mean square error (MSE),
- problems:
① the first question emerges: can we design CNN architectures that directly implement LR as input to address these problems
② how can we design such models of greater complexity?
B. State-of-the-Art Deep SISR Networks
Learning Effective Upsampling with CNN:
- upsampling operation：deconvolution [50] or transposed convolution [51]
Given the upsampling factor, the deconvolution layer is composed of an arbitrary interpolation operator (usually, we choose the nearest neighbor interpolation for simplicity) and a following convolution operator with a stride of 1, as shown in Fig. 3.（近邻插值，乘卷积核）
