Research On Image Super-resolution Reconstruction Based On Convolutional Neural Network

As a carrier of information transmission,images play an indispensable role in daily life.The quality of the image to be processed directly determines the efficiency and difficulty of completing image processing tasks.Image resolution characterizes the ability of an image or display device to express texture information,high-resolution images have more details.Affected by the cost of hardware equipment and the changing acquisition environment,the quality of some images cannot meet the requirements of image processing tasks.Therefore,the use of software algorithms to reconstruct super-resolution has become one of the current research hotspots in the field of computer vision and image processing.This technology is dedicated to efficiently and quickly reconstruct high-resolution images from low-resolution images.With the rapid development of artificial intelligence technology and the improvement of computer performance,super-resolution reconstruction algorithms based on convolutional neural networks came into being.Many traditional methods,including neighborhood embedding and sparse coding,make full use of theoretical knowledge and show strong performance in the field of super-resolution reconstruction,but traditional methods cannot effectively use standard data sets to simulate image degradation and reconstruction processes,and deep learning methods that make full use of prior knowledge have become mainstream.The main work of this article is as follows:(1)In response to existing methods tending to use deeper and deeper networks,resulting in problems such as excessive computational cost,loss of shallow information and long training time,a cascading residual network is used to quickly propagate low-level features to high-level through multi-level representation and fast connection.At the same time,the residual network is improved to remove batch normalization and save computing resources,and use a smaller convolution kernel to expand the number of channels to allow more low-level information to pass.Finally,low-rank convolution is used to extract features and reduce parameters.On the premise of improving the training speed,the quality of the reconstructed picture is guaranteed.(2)The existing super-resolution reconstruction methods based on deep learning are basically supervised,use standard data sets,and train under known degradation factors.It may cause the image to be too smooth,lack texture details,and the training effect is not good when the imaging conditions are unknown.This paper proposes an unsupervised super-resolution method,using the characteristic that the internal image cross-entropy loss is smaller than the external image,and only using a simple convolutional neural network to train the input image itself,not need to use deep network to pre-train standard data set.Aiming at the problem of single loss function pursuing data results and ignoring human eye perception,a compound loss function is used to take into account both visual effects and data evaluation.In the case of limited computing resources consumption,this model has achieved very good results.