Light cancellation for hyperspectral material classification

This thesis work addresses the problem of hyperspectral reconstruction and illumination estimation to enhance material classification capabilities. Hyperspectral images (HSIs) store detailed information across multiple, continuous spectral bands. Offering richer information compared to standard RGB images, they are particularly suitable in vision tasks such as object classification and recognition. To acquire spectral images in a dynamic and portable way, single-shot imaging systems, such as the Multispectral Color Filter Array (MSFA), have been developed. The MFSA consists of a set of color filters (8 in ours) into a CMOS image sensor, producing a single-channel mosaicked image where each pixel records information from one spectral channel. The use of single-shot techniques introduces the need of reconstruction algorithms able to recover the full HSI. This reconstruction is crucial for deriving the spectral reflectance of objects, an intrinsic property of materials that aids in their classification. Accurate spectral reflectance extraction depends on precise knowledge of the illumination conditions, which is often difficult to achieve. Therefore, estimating the scene's illumination is a practical approach to obtaining reflectance data. The reconstructed data is used for material classification, with a focus on minimizing light interference to improve accuracy. Through experimentation, we demonstrate that our light cancellation method significantly enhances material classification precision. The results suggest that incorporating light cancellation in hyperspectral imaging systems is a promising approach for advancing material detection and identification technologies.