Research On Image Super Resolution Algorithm Based On Deep Learning

The image super-resolution reconstruction technique can reconstruct a high resolution image from only a single low-resolution image.With the continuous development of artificial intelligence,super-resolution reconstruction technique has been widely applied and researched in the fields of video and image compression and transmission,medical imaging,remote sensing imaging,video sensing and surveillance.Therefore,the single image super-resolution has been a hot research direction for computer vision and image processing fields.At present,the image super-resolution methods can be mainly divided into the following three categories: the method based on image interpolation,the method based on image reconstruction,and the method based on learning.The superresolution method based on deep learning can make full use of the feature information of the image,directly learn the relationship between the low-resolution image and the high-resolution image by constructing a model,and then use the trained model to reconstruct the high-resolution image block.Deep learning-based methods can significantly improve the quality of single-image super-resolution.This thesis deeply studies the reconstruction algorithm of deep learning in the field of image superresolution,and finds that the current research tends to use deeper convolutional neural networks to enhance network performance.But blindly stacking a single convolutional layer to increase the depth of the network cannot effectively improve the network and improve performance.As the depth of the network increases,more problems will appear during the training process,and more training skills and greater computing power are required.It ignores the diversity and complexity of image features and the full use of features.In order to solve the above problems,this paper improves on the basis of residual neural network,and proposes two algorithms basedon adaptive fusion residual network and two-scale fusion residual network,which are applied to the task of image super-resolution.The contribution points of the thesis are expressed in the following:(1)Most existing image super-resolution models suffer from insufficient utilization of image features.In order to solve the above problems,this paper improves on the basis of the residual network and proposes an image super-resolution algorithm based on the adaptive fusion residual network.The algorithm constructs an adaptive fusion residual block,which is used to adaptively select valuable information in multi-layer features for fusion.The algorithm also simplifies commonly used reconstruction modules,making it simpler and more flexible.According to the connection between the multi-layer image features,the adaptive fusion residual block can extract and fuse the valuable information in the multi-layer image features,and use this information to obtain high-quality images.In the task of image superresolution,the image reconstruction performance of this algorithm model is much improved compared with many algorithms.(2)Most existing image super-resolution algorithms blindly increase the depth of the network to enhance the performance of the network,but ignore the full use of image features of different scales.In order to solve the above problems,this paper improves based on the residual network,and proposes an image super-resolution algorithm based on the two-scale fusion residual network.The algorithm constructs a dual-scale fusion residual block for adaptively detecting image features of different scales.Based on the two-scale fusion residual block,a feature adaptive transfer structure is also constructed,which is used to extract valuable image features and transfer them to the end of the network for global fusion to improve the reconstruction performance of the model.The model of this algorithm can make full use of the valuable image features of different scales to obtain high-quality images.It achieves good objective evaluation and visual effects in the tasks of image super-resolution.