Fast feature selection algorithm for poultry skin tumor detection in hyperspectral data

Hyperspectral reflectance imaging data are analyzed for poultry skin tumor detection. We consider selecting only a few wavebands from hyperspectral data for potential use in a real-time multispectral camera. To do this, we improve our prior tumor detection system by employing our new adaptive branch and bound algorithm and a support vector machine classifier. Our HS analysis is useful since it provides a guideline for selection of the specific wavelengths for best tumor detection (feature selection). Experimental results demonstrate that our optimal adaptive branch and bound algorithm is significantly faster than other versions of the branch and bound algorithm. We compare the performance of our feature selection algorithm to that of a feature extraction algorithm and show that using our feature selection algorithm gives a better tumor detection rate and a lower false alarm rate.