Super-resolution Image Reconstruction Based On Information Distillation And Attention Mechanism

Image is the medium of human cognition,and high-resolution image has become an important developing direction of color and spectral imaging equipment with abundant spatial structure and detail information.Image super-resolution technology comes into being,that is,through a series of technical means to improve the resolution and quality of the image.With the rise of 5G technology and deep learning,image super-resolution technology has also made positive progress in theory and application.Most of the current super-resolution algorithms rely on deep and complex network models to achieve higher performance,but it also leads to large model parameters and high computational complexity.In addition,the existing neural network-based image super-resolution methods do not sufficiently extract and reuse features,which leads to poor fitting ability of the model.In order to solve the problems of high computational complexity and low feature reusability in image super-resolution reconstruction,the real scene image can be reconstructed better.Based on the existing research,this paper improves the information distillation module and spatial attention mechanism in information distillation network,and proposes an image super-resolution reconstruction method based on information distillation and attention mechanism.The rationality of the network layer is verified by experiments.The low-level feature information and highlevel feature information are fully utilized while the complexity of the model is reduced.In view of the low computational complexity and good reconstruction effect of IDN(Information Distillation Network)Network,this paper proposes an IDN-p(Distillation Network Plus)Network by using the partial structure of its Information Distillation block and improving it.In this paper,the model network framework is post-sampling,in which the primary features are extracted in the feature extraction module,and the advanced features are extracted and the weights are assigned in the multi-feature fusion module In the part of image prediction,sub-pixel convolution is used for super-resolution image reconstruction.In order to reduce the degradation of the model caused by the increase of network layers,the structure of embedding residuals in the information distillation group is adopted in the multi-feature fusion module.A local residual unit is added to the information distillation group to avoid the loss of features and reduce the difficulty of training network.The low-level feature and high-level feature information are fully utilized to improve the reconstruction effect through long-short path jumping connection.In addition,the model introduced spatial attention module after all the information distillation group to refine the advanced features.The global features after fusion are weighted,and the key features are given high weight,while the non-key features are penalized.In this paper,ablation experiments are designed to verify the rationality of the model structure,and hyperparametric experiments are designed to verify the rationality of the model layers.The objective index of the image super-resolution reconstruction method is used to evaluate the reconstruction effect of the proposed model on the selected test set.Compared with the image super-resolution classical method in different scenes,it is proved that the model designed in this paper can display good reconstruction ability on subjective and objective level,the local key information and texture details of the reconstructed high-resolution image are displayed.