Effect analysis of Self built low-quality scene image set super-resolution reconstruction based on deep learning technology
Ying Yan, Liu KeXin · 2024
High quality images contain richer information. However, often due to the inherent limitations of hardware devices, the resulting image resolution is low and critical information is lost. In order to improve image resolution, SRCNN(the Super Resolution Convolutional Neural Networks) and FSRCNN(the Fast Super Resolution Convolutional Neural Networks) models based on deep learning techniques have achieved better reconstruction results compared to traditional super-resolution methods. But currently, most of the experimental objects for testing the performance of these algorithms focus on low-quality images after artificial interference, with little consideration given to low-quality images captured in complex real-world scenarios. To address the aforementioned issues, this article first establishes three types of low-quality image scene sets: visible light difference, harsh climate, and machine performance limitations. Secondly, comparative experiments are conducted by setting different parameters and using different models. Finally, suitable models for reconstructing different scene sets are analyzed and found.