Improving spatial quality of terrestrial and satellite images by demosaicking and fusion
University of New Brunswick
Improving the spatial quality of a colour image brings valuable benefits to all imaginable applications of the image. One method for such an improvement is to incorporate ‘panchromatic’ sensors into imaging. Panchromatic sensors provide images with higher spatial quality than colour images because they do not filter any complementary colours of the incoming light. Combining panchromatic and colour sensors has been employed in different fields. In remote sensing (RS), panchromatic and multispectral / hyperspectral images are captured by two separate sensor chips and then fused through pan-sharpening techniques. In terrestrial applications, a single sensor chip is used to accommodate both panchromatic (or white, W) and colour (RGB) pixels using a Colour Filter Array (RGBW CFA). A ‘demosaicking’ procedure needs to be employed to generate RGB colour images. Both pan-sharpening and RGBW demosaicking still have unsolved problems despite being used by the imaging industry for a while. In RS, most hyperspectral bands are not pan-sharpened, because they fall beyond the panchromatic spectral range, causing significant spectral distortion. For RGBW demosaicking, limited methods have been published which produce low-quality images, mainly because they demosaick panchromatic and colour images independently. Another issue is that existing approaches cannot handle images corrupted by noise, because they do not involve denoising. This dissertation aims to overcome the above-mentioned obstacles in improving the spatial quality of the hyperspectral/colour images. For hyperspectral images, this research develops an adaptive band selection strategy to identify the bands across the entire spectrum that can be pan-sharpened without introducing high spectral distortion. For RGBW demosaicking, this research firstly proposes a collaborative interpolation between panchromatic and colour pixels. It significantly improves the spatial quality by reducing the zipper effects and retaining the spatial details. The research then proposes a deep learning-based approach for RGBW joint demosaicking and denoising, along with a procedure to prepare the required training dataset. Results show a considerable quality improvement over existing methods even for images corrupted by various noise levels. In summary, this research leads to improving the spatial quality of those hyperspectral bands, that were previously left unfused. It also increases the potential of using RGBW cameras in daily applications due to the significant quality boost.