Image Super Resolution Reconstruction Based On Feature Region Constraint Learning

Image super-resolution reconstruction is not only an important area in image processing,but also is one of the hot research directions of computer vision currently.The core of its technology is to reconstruct a high-resolution image from one or more low-resolution images.In most computer vision processing tasks,high-resolution images are often needed for its subsequent image processing,analysis,and understanding.Especially in the fields of security surveillance,it is critical to improve the image quality since the surveillance videos are commonly highly compressed.However,due to the influence of image acquisition equipment and environment,the acquired image data may not meet quality requirement for the computer vision task.Conventionally it is costly and time-consuming by improving hardware to increase image resolution.Therefore,improving the resolution of surveillance image data through software algorithms has become a feasible solution.At present,there are mainly methods based on interpolation,methods based on prior knowledge,and methods based on deep learning.The first two categories have shortcomings such as block effect and poor performance when applied in a large scale.Deep learning method has one significant advantage on automatically extracting image features compared to other approaches.However,it is difficult to obtain high similarity features from deep features for super-resolution reconstruction of the image.Therefore,this paper mainly focuses on the learning-based image super-resolution reconstruction technology,and designs a feature region constraint learning module to exploit similar features between feature maps for super-resolution reconstruction.Its main work is as follows:We review and study the basic principles and model design methods of feature extraction by convolutional neural networks.Several classic convolutional neural network algorithms,such as LeNet,AlexNet,VGG,ResNet,etc.,have been studied and analyzed.The advantages of convolutional neural networks for extracting features in different scale spaces are verified,and it is found that there are regionally similar features between feature maps of different scales.Therefore,this paper proposes an approach that uses similar feature constraint learning between feature maps to improve image super-resolution reconstruction.Based on the deep convolutional neural network and residual learning module,a similar feature constraint learning module is added,and the loss function is optimized to improve the feature extraction and refinement to improve reconstruction quality.In terms of experiments,this paper conducts experiments on the standard DIV2K high-definition dataset in natural scenes and the surveillance image dataset by our cameras.The experiments demonstrate the model's superiority and robustness for both image categories.Finally,this paper studies several classic image super-resolution algorithms based on convolutional neural networks,such as SRCNN,EDSR,SRGAN,DBPN,etc.,and conducts various experiments to make comparisons with our algorithm proposed in this paper.