Research On The Methods Of Deep Convolutional Neural Network Based Single Image Super-Resolution

Single image super-resolution(SISR)reconstruction is an important task in the field of artificial intelligence and computer vision,which is deeply respected by academia and industry since 1970s.It can be summed up that recover the high-resolution image from a given low-resolution image input.Compared with low-resolution images,high-resolution images usually contained larger pixel density,richer texture details and higher degree of reliability.SISR plays an important role in many field such as Image Compression,Medical Imaging,Remotely-sensed Image,Security field and Video Perception.However,SISR still faced some challenges because the low-resolution image inputs drop numerous high frequency details,which results in the difficulty of recovering immediately.Recently,deep learning has shown its incredible capacity in computer vision.Therefore,research on the methods of deep convolution neural network based single image super-resolution is a very important but challenging task.The main works of the thesis are introduced as follows.Firstly,we introduce the basic concepts and parts of the deep convolution neural network in detail.Secondly,we further illustrate the developmeit history of SISR from traditional methods to deep learning based on methods.What' s more,we also analyze the characteristics of these methods.Besides,we proposed a residual inception module based CNN for SISR.With the development of ResNet,CNN has alleviated the gradient disappear problem effectively.Consequently,many SISR methods based on CNN design deeper and more parameters structure,which bring the better performance but training difficulty.While residual inception module based CNN focused on designing the efficient feature extraction module:use the different size of convolutional kernel to extract the abundant image feature and cascade theqm for fusion on the feature channel.At the same time,this method apply three residual structure:residual in the inception module,local residual and global residual.The application of theses residual structures improve the model',s capacity and make the model easier to train.Furthermore,we proposed a skipping residual inception module based CNN for SISR.This method focused on the whole network structure:compress the residual inception module and stack more modules for deepening the network.Meanwhile,it introduced the dense and skipping mode to connect the lightweight residual inception module.Benefit from this connection mode,our model strengthened the delivery of feature information powerfully,reuseed the shallow features directly,boosted the performance extremely.Compare with the residual inception module based CNN,the skipping residual inception module based CNN has deeper structure and better performance but more training time and lower speed of reconstruction.They focus on different eyes and have their own advantages.Last but not least,we show our experiment design from all aspects in detail.Extensive evaluation on benchmark datasets show that the proposed model achieves good performance against state-of-the-art methods.