Research On Lightweight Image Super-Resolution Reconstruction Based On Deep Neural Network

Image resolution is a measure of the richness of detail in an image and reflects the ability of a physical imaging system to capture the spatial information of a scene,with higher resolution images containing more detailed textures,greater pixel density and higher reliability.In practice,physical imaging systems are limited by optical sensor equipment,complex imaging environments,transmission media and bandwidth,making it difficult to directly capture high-quality,high-resolution images with rich detail and texture.In many practical applications,such as medical imaging,remote sensing imaging,traffic security,etc.,improving the spatial resolution of images plays a crucial role.Image Super-Resolution can be used to improve image resolution and image quality without changing the existing hardware,thus acquiring result is high-quality,high-resolution images that meet the needs of downstream scientific analysis and practical applications.Recently,there have been significant breakthroughs and advances in SR reconstruction of natural images,benefiting from deep learning techniques and neural network models,with a variety of outstanding State-Of-The-Art(SOTA)models springing up.Currently,driven by the paradigm of big data and big models,the computational complexity and number of parameters of image SR models are increasing day by day,making the deployment and application of the models in edge devices limited.To address this problem,this thesis focuses on lightweight image SR reconstruction techniques based on deep neural networks,and proposes two new lightweight SR reconstruction models.The main studied are summarized as follows.1.A lightweight image super-resolution network based on hierarchical feature reconstruction.Out of the thinking and understanding of image restoration,a basic premise assumption is derived that images can be decomposed into different signal components,i.e.images can be considered as a synthesis of different feature components.Therefore,based on the idea of Disentangled Representation Learning,a lightweight image SR network(HFRN-SR)based on hierarchical image feature reconstruction is proposed.Firstly,in order to ensure the content consistency between the reconstructed image and the original image,the result of bicubic interpolation is used for the low-resolution image as a component to retain the lowfrequency energy and color information.Then,a lightweight Adaptive Inverse Residual Block is constructed in order to extract the shallow coarse feature components of the image reconstruction.A Refined Feature Module is designed in order to obtain refined details and finer-grained feature components.Finally,the interpolated image,shallow coarse features and refined features are integrated to obtain the final SR reconstruction results.Extensive experiments and model analysis show that the proposed HFRN-SR outperforms some SOTA lightweight SR models in terms of both quantitative and qualitative performance.2.A lightweight image super-resolution network based on asymmetric encoder-decoder.Memory and computational resources of edge devices are limited,so that the deployment and application of deep neural networks on embedded and mobile devices face great challenges.To solve this problem,this thesis proposes a lightweight SR model(LSRN-AED)with the core idea of compromising SR model complexity and performance,i.e.,achieving better performance while reducing SR model parameters.1)In rethinking the role of encoder and decoder in image recovery,an Asymmetric Encoder-Decoder(AED)consisting of a complex encoder and a simple decoder is designed to achieve image feature extraction and reconstruction.2)A simple structure is built for the image feature reconstruction task undertaken by the decoder,a simple structured inverse residual block is constructed to reduce the computational cost of the model and the mapping of redundant features.3)Inspired by the Transformer structure,the encoder is designed as an Epiphany Encoder to implement image feature extraction and characterization for better compression and coding of image signals.The final experimental design and analysis show that the proposed LSRN-AED with multiple AEDs not only reduces the complexity of the SR model,but also achieves better SR reconstruction performance,which is better than the existing SOTA lightweight SR model.