Article reproduction: srcnn

12 year ,AlexNet stay ImageNet Super performance in image classification competition （ Reference link 【1】）, The powerful ability of convolutional neural network in image classification is proved , Later, various scholars in CNN A variety of improved image classification network architectures are proposed .

14 year ,DONG For the first time CNN To carry out super-resolution image reconstruction （ Reference link 【2】）, The article puts forward SRCNN The architecture of . Image hyperfractionation can be considered as a pixel level regression task .

stay climate field , Generally, it is a return mission , After all, most of them are continuous values .

Let's reproduce it in this article SRCNN.

Article summary

We propose a deep learning method for single image super- resolution (SR). Our method directly learns an end-to-end mapping be- tween the low/high-resolution images. The mapping is represented as a deep convolutional neural network (CNN) [15] that takes the low- resolution image as the input and outputs the high-resolution one. We further show that traditional sparse-coding-based SR methods can also be viewed as a deep convolutional network. But unlike traditional meth- ods that handle each component separately, our method jointly optimizes all layers. Our deep CNN has a lightweight structure, yet demonstrates state-of-the-art restoration quality, and achieves fast speed for practical on-line usage.

The general idea is that this paper proposes a lightweight 、 End to end 、 be based on CNN The neural network of , It outperforms the traditional sparse coding method in the task of image super segmentation , And it's very fast .

Network architecture

SRCNN The network architecture of is really simple , Its network has only three convolution layers ：

SRCNN- Quoted from reference 【2】

The corresponding code ：

class SRCNN(nn.Module):
    def __init__(self, inchannels):
        super(SRCNN, self).__init__()

        self.main = nn.Sequential(
            nn.Conv2d(inchannels, 64, kernel_size=9, stride=(1, 1), padding=(4, 4)),
            nn.ReLU(),

            nn.Conv2d(64, 32, kernel_size=1, stride=(1, 1), padding=(0, 0)),
            nn.ReLU(),

            nn.Conv2d(32, 1, kernel_size=5, stride=(1, 1), padding=(2, 2))
        )

    def forward(self, x):
        y = self.main(x)   
        return y

• This article uses 9-1-5 Convolution kernel , This can be adjusted by yourself ;
• SRCNN The input of is to use bilinear interpolation to the target size , Then input it into the network ;
• Each layer padding The quantity of can be calculated by yourself , Make sure the input and output are the same size ;

SRCNN It's the first chapter of using depth convolution to carry out hyperdivision , It is also a groundbreaking work , But I want to apply it according to my own research , It can only be said that the resistance is long ....

Reference resources

【1】CNN The first article for image classification ：KRIZHEVSKY A, SUTSKEVER I, HINTON G E J C O T A 2012. ImageNet classification with deep convolutional neural networks. 60: 84 - 90.
【2】CNN The first article for image regression ：DONG C, LOY C C, HE K, et al. Learning a Deep Convolutional Network for Image Super-Resolution[C]//Computer Vision – ECCV 2014.Springer International Publishing,2014:184-199. 10.1007/978-3-319-10593-2_13.