Research On Lightweight Image Super-Resolution Methods With Multi-Scale Feature Extraction

High resolution images have better imaging quality and clear texture features,which can not only meet the perceptual needs of human eyes,but also better help to complete advanced computer vision tasks.Image super-resolution is a technology to transfer low resolution images to high resolution images by using some algorithms.In recent years,the application of deep convolution neural network has greatly promoted the development of computer vision,and the technology of image super-resolution has also been greatly improved.Although the current models based on deep convolution neural network have achieved a great breakthrough compared with the previous traditional algorithms,most of these models have a large number of parameters and complex calculation and can not be well used in some lightweight devices.Therefore,lightweight image super-resolution is a more challenging and pratical problem.The features of different scale play an important role in the performance of image reconstruction.Therefore,this thesis combines multi-scale feature extraction to deal with the problem of lightweight image super-resolution.And,this thesis carries out the following tasks:(1)In this thesis,a lightweight multi-scale feature guided image super-resolution network was proposed.Firstly,in order to extract the features of different receptive fields,a multi-scale feature extraction block was designed to improve the feature extraction ability of the model.Secondly,this thesis proposed a feature guided connection mechanism to make use of different levels of features.Low-level features were used for feature extraction and high-level features were used for image reconstruction.Compared with dense connection mechanism,it had less parameters and computation.At the same time,a combination of attention fusion block and feature guidance connection was proposed to distribute the weight of high-level features,so that the model could pay attention to the information of different dimensions of high-level features,so as to fuse high-level features for image reconstruction.(2)In this thesis,a lightweight multi-scale hierarchical feature fusion image superresolution network was proposed.Specifically,the multi-scale attention module was used to extract the feature information at different scales,and the features of different scales were extracted by means of hierarchical feature concattenation and hierarchical residual connection,which improved the feature extraction ability of the network.In addition,the hierarchical feature fusion block was used to fuse the features of different levels extracted from the network,which maked full use of the feature information of different levels,and improved the reconstruction performance of the network by hierarchical fusion of high-level and low-level features.The two methods use different ways to construct multi-scale structures and use different levels of features.In comparison,the first method has better performance and the second method is more lightweight.On the test data set,compared with the current mainstream lightweight image super-resolution methods,our two methods have a certain improvement in the objective evaluation index and visual quality,which proves the effectiveness of the two methods proposed in this thesis,and further proves the effectiveness of our methods in the ablation experiment and analysis of model components.Our two methods have certain research value in lightweight image super-resolution.