Demosaicing of Multispectral Imaging Camera data using Neural Networks
Sumita Gupta, Sapna Gambhir, Saru Dhir, Simran Sapui, Rana Majumdar, Deepsikha Naskar · 2024
In most digital cameras, the sensor uses a Bayer filter array to capture the image. This array records only one colour, either blue or green or red for every image pixel, resulting in a mosaic image. To retrieve the missing colour information in the captured image making use of cross-channel interpolation is referred to as demosaicing. Demosaicing, especially in the context of colour demosaicing (CDM), plays a vital role as the first step in obtaining high-quality colour images with single-chip cameras. Traditional demosaicing methods, used to reconstruct colour images from raw sensor data in digital cameras, have several drawbacks. They often result in a loss of spatial resolution, introduce artifacts, lack robustness in challenging conditions, and may require manual tuning for optimal performance. In contrast, neural networks offer advantages like end-to-end learning, improved image quality, robustness, flexibility, and state-of-the-art performance. They learn complex mappings directly from raw data, adapt to various conditions, and produce visually pleasing, high-resolution images. However, they require substantial training data and computational resources, making them less suitable for resource-constrained applications. In this paper, conventional interpolation methods are compared to deep-learning based approaches for image demosaicing. To validate the approach to image demosaicing, aerial images captured from the Mars Colour Camera have been used.