A major problem for obtaining target reflectance via hyperspectral imaging systems is the presence of illumination and shadow effects. These factors are common artefacts, especially when dealing with a hyperspectral imaging system that has sensors in the visible to near infrared region. This region is known to have highly scattered and diffuse radiance which can modify the energy recorded by the imaging system. Shadow effect will lower the target reflectance values due to the small radiant energy impinging on the target surface. Combined with illumination artefacts, such as diffuse scattering from the surrounding targets, background or environment, the shape of the shadowed target reflectance will be altered. In this study we propose a new method to compensate for illumination and shadow effects on hyperspectral imageries by using a polarization technique. This technique, called spectropolarimetry, estimates the direct and diffuse irradiance based on two images, taken with and without a polarizer. The method is evaluated using a spectral 2 similarity measure, angle and distance metric. The results of indoor and outdoor tests have shown that using the spectro-polarimetry technique can improve the spectral constancy between shadow and full illumination spectra.

No abstract

No abstract

This paper reports an extension of the previous MIT and Caltech's cortex-like machine vision models of Graph-Based Visual Saliency (GBVS) and Feature Hierarchy Library (FHLIB), to remedy some of the undesirable drawbacks in these early models which improve object recognition efficiency. Enhancements in three areas, a) extraction of features from the most salient region of interest (ROI) and their rearrangement in a ranked manner, rather than random extraction over the whole image as in the previous models, b) exploitation of larger patches in the C1 and S2 layers to improve spatial resolutions, c) a more versatile template matching mechanism without the need of 'pre-storing' physical locations of features as in previous models, have been the main contributions of the present work. The improved model is validated using 3 different types of datasets which shows an average of ~7% better recognition accuracy over the original FHLIB model.