An Improved Approach for Small Object Detection in Hyperspectral Images

Abstract

Due to the fact hyperspectral cameras have low spatial resolution values, small target detection becomes a challenging task. In this study, a new method was proposed to detect small targets with high performance values. For target detection algorithms, it is very important to extract the accurate statistical informations of the image. In particular, accurate background information is very important for the Generalized Likelihood Ratio Test (GLRT). In order to extract these statistics correctly, the number of pixels of the image should not be too many or too few. For this reason, the hyperspectral image passed through the preprocessing steps and the image is divided into small tiles depending on the target dimensions to be detected. The target detection algorithm is performed separately on each of the tile components. In this way, the number of pixels from which the background information of the image is extracted is limited. Then, the target detection results obtained from the small pieces are combined and a general result map is obtained. The tests were performed on 3 different targets in 2 different images. When the results were evaluated, it was observed that the detection performance values obtained using the proposed method were higher than the detection performance values obtained using the GLRT algorithm on the whole image.