Research On Image Super Resolution Reconstruction Based On Convolutional Neural Network

Image is a important carrier in the process of obtaining information by humans through vision.Because of its intuitive and rapid characteristics,it plays an indispensable role in the process of information transmission.70% of the content of human object information comes from images The higher the sharpness of the image,the richer the detailed information is,which is more conducive to the post-processing of the image.Image super-resolution reconstruction technology generally refers to the process of using computer vision or digital image processing technology to reconstruct a high-resolution image from one or more low-resolution images.With the rapid development of artificial intelligence in recent years,compared with traditional algorithms,image super-resolution based on convolutional neural networks is gradually becoming the mainstream method due to its good effect of end-to-end learning.Aiming at the shortcomings of the existing convolutional neural network-based models,this paper mainly researches from the following two aspects:(1)Image super-resolution reconstruction based on Depth jumping cascade: For the current network models based on convolutional neural network models,the convergence rate is slow,the original images need to be pre-processed before training,and the redundancy in the network,etc.In this paper,a method for super-resolution reconstruction of a single image based on depth hopping cascade is proposed.First of all,the algorithm omits preprocessing and extracts shallow features directly on low-resolution images.Secondly,by using skip cascading blocks,the features of each image can be fully extracted from each convolutional layer to achieve feature reuse and reduce network redundancy.Sex.The jumping concatenated block of the network can directly establish a short connection from the output to each layer,speed up the network's convergence speed,alleviate the problem of gradient disappearance,and finally use sub-pixel convolution to enlarge the image.Experiments show that on several public data sets,both from the objective evaluation index and subjective effect of the algorithm,they are better than the existing algorithms,which fully demonstrates the excellent performance of the algorithm.(2)Image super-resolution reconstruction based on light recursive residuals: For existing models due to the excessive number of network layers,the amount of model parameters and computational complexity increase and the speed is too slow,which is not suitable for application to embedded and mobile devices And other problems,a light recursive residual image super-resolution reconstruction network model is proposed.Residual errors are introduced to alleviate the problem of gradient disappearance and network degradation in deep networks.Recursive structure and group convolution are used to reduce the amount of calculation.Reduce the complexity of the model,and the model does not need to preprocess the image.Experimental results show that the proposed algorithm reduces the amount of parameters and complexity of the model while ensuring the quality of image reconstruction,and the reconstruction speed is also improved to a certain extent.