Subsurface Material Estimation Using Hyperspectral Imaging
- Publication Type:
- Thesis
- Issue Date:
- 2022
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Traditional vision is focused on the perception of visible light but limits us to only detecting surface features. Near-infrared light can penetrate certain materials, so to capture this light, hyperspectral cameras can be employed. However, the images from these cameras lack spatial understanding. This work centres on a multi-modal vision approach that leverages the capabilities of modalities such as colour and depth to improve the estimation of material properties, which can be used to detect features within.
Camera calibration is required for any camera to remove geometric discrepancies in images and estimate camera poses. Generally, it involves an optimisation that uses images of a known pattern in various poses. This task is challenging for hyperspectral cameras as they typically capture images in a line-scan manner, with only a single spatial dimension. To aid with calibration, a multi-modal camera system is deployed by combining the hyperspectral with an RGB-D camera, and an active calibration algorithm is devised, which selectively chooses the best images to improve the calibration verified through experiments.
Illumination from external light sources is essential to image formation, especially for hyperspectral cameras. Real light sources have an asymmetric distribution in intensity but are modelled as symmetric. Therefore, the second contribution of this thesis is spatially modelling the distribution of a light source through a novel data-driven Gaussian process (GP) model. The intensity distribution that is estimated by the proposed GP model shows less error while capturing the inherent asymmetries.
Cameras measure radiance that depends on illumination, shape and the respective material. Reflectance, an intrinsic property of a material, is independent of these factors. The reflectance can be estimated from the radiance by assuming that the light interactions with a given material follow the dichromatic reflectance model. The third contribution of this thesis involved the recovery of reflectance using prior modelling of light sources and surface shape. Evaluation in recovering reflectance shows the least error using the proposed reflectance estimation method with less variation across areas containing the same material.
Finally, a case study is investigated to estimate subcutaneous fat depth on lamb carcasses using hyperspectral imaging. Ground truth fat depth was acquired from a CT scanner, and the previous contributions were used to estimate the reflectance of the cuts. Results across all regression fits for each reflectance estimation method showed that the fat depth was best modelled using deep learning methods with the proposed estimated reflectance method.
Please use this identifier to cite or link to this item: