SH-RNet: Dual-branch Network for Joint Specular Highlight and Reflection Removal

Removing reflections from an image is a very important task in image processing. Many times, we capture photographs of objects inside a glass. In such cases, the image contains glass reflections, which can sometimes be intense. There are many state-of-the-art methods for reflection removal. Although they produce good results, they cannot effectively remove intense reflections. This paper focuses on removing intense reflections from an image. Most state-of-the-art methods focus only on glass reflections. However, specular highlights, another important component, commonly occur as part of reflections. Specular highlights are strong, shiny spots that degrade image quality. Reflections and specular highlights are image degradation artifacts that adversely affect visual quality and reduce the performance of computer vision applications. To address this problem, a Specular Highlights and Reflection Removal Network (SH-RNet) has been proposed. The algorithm removes both glass reflections and specular highlights. The Single Image Reflection Removal Dataset, a benchmark dataset, is used to conduct a range of experiments. The proposed model demonstrates significant improvement over existing state-of-the-art techniques. A lightweight deep learning model is employed that performs better when images contain intense reflections. The results indicate that the proposed framework effectively suppresses reflections and specular highlights while preserving image details, making it suitable for applications in image enhancement, object recognition, and computer vision systems.