Image quality improvement on surveillance footage using super-resolution techniques to generate composite images

Abstract

Super-resolution is the process of enhancing the resolution of a captured image through the process of upscaling and enhancing a lower-resolution image to construct a higher-resolution image. In recent years, we have witnessed several advancements in the state-of-the-art Deep Learning-based architectures, which also includes remarkable progress in the task of image super-resolution. With many possible interesting applications of super-resolution, we have identified a potential application of super-resolution that could be useful in the field of forensic science, particularly in the area of surveillance. This project was designed to generate a composite image of the masked face of a suspect captured on a surveillance camera. We have assembled three models that were essentially the baseline model equipped with super-resolution layers for enhancement on the input image and/or on the baseline model’s output image. Our models were compared against the baseline model, which is an image inpainting model that does not have any super-resolution layers. All the three proposed models yielded satisfactory results, and one of the proposed models exceeded the performance of the others in terms of visual quality. This project has successfully demonstrated that the application of super-resolution technique can help in the image quality improvement, which, therefore, will help in generating a more accurate composite picture.