Recent Advances in Image Super-Resolution Reconstruction Based on Machine Learning

Image super-resolution (SR) reconstruction is a crucial research area in computer vision, aiming to restore high-resolution images from low-resolution inputs, thereby enhancing image detail and quality. With the continuous growth of digital image applications, SR technology has been widely utilized inAfields such as medical imaging, satellite remote sensing, surveillance video enhancement, and virtual reality. However, despite significant progress in objective image quality, existing SR methods still face challenges such as loss of image details, unnatural textures, and visual inconsistencies. This is especially evident in complex scenes or high-noise environments, where traditional unified models are ineffective in addressing the differences between image regions, resulting in suboptimal reconstruction outcomes. In recent years, deep learning methods, such as Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), have made remarkable strides in the field of SR. However, most methods still overlook the spatial dependencies between different regions of the image. To address this limitation, this paper proposes a SR reconstruction framework based on the Markov Decision Process (MDP) and Deep QNetworks (DQN), which dynamically selects SR models using reinforcement learning principles for adaptive optimization across image regions. Furthermore, aAnew reward function is introduced to resolve the model selection consistency issue across regions, aiming to improve the visual transition between adjacent regions and enhance the overall perceptual quality of the image. Experimental results demonstrate that the proposed framework effectively improves the reconstruction performance of SR images, significantly enhancing visual coherence while maintaining objective quality.