Image super-resolution reconstruction method based on improved attention mechanism

Image super-resolution enhancement technology occupies a central position in visual information processing in the digital era, and its core task is to reconstruct high-resolution images using low-resolution image data. With the development of computing technology, image super-resolution techniques are not only theoretically significant, but also become more and more critical in practical applications, and it has become a hot area of computer vision research. Nevertheless, there are still some challenges in current super-resolution solutions. Existing models are deficient in extracting the diversity of image features, which leads to the problems of detail loss and excessive smoothing of the reconstructed images, which seriously affects the visual quality of the images and makes the reconstructed images visually lacking in realism. In order to overcome these difficulties, this study focuses on the integration of image multi as features as well as exploring the attention mechanism of arbitrary scale images. Through these studies, we aim to improve the performance and efficiency of super-resolution reconstruction techniques. The main contributions of this thesis include:Aiming at the limitations of the multi-feature self-attention mechanism, this study proposes several innovative attention operators. These operators not only enhance the model’s feature extraction ability to capture image details and texture information more efficiently, but also improve the diversity of feature information. In addition, in order to overcome the limitations of existing attention mechanisms in feature extraction, this study introduces a novel, parameter-free attention module. This module effectively integrates information from both channel and spatial dimensions, significantly enhancing the feature representation. This design simplifies the model structure, avoids complex operations, and enhances model performance in image processing and computer vision tasks.