Target detection in multispectral images using the spectral co-occurrence matrix and entropy thresholding

Relative entropy thresholding techniques have been used for segmentation of objects from background in gray-level images. These techniques are related to entropy-based segmentations computed for the statistics of a spatial co-occurance matrix. For detection of spectrally active targets such as chemical vapor clouds in multispectral or hyperspectral imagery, a spectral co-occurrence matrix is employed. Using the entropy of various regions of the matrix, thresholds can be derived that will segment an image family based on the spectral characteristics of the intended target. Experiments are presented that show the detection of a chemical vapor cloud in multispectral thermal imagery. Several manners of dividing the co-occurance matrix into regions are explored. Thresholds are determined on both a local and global basis and compared. Locally generated thresholds are treated as a distribution and separated into classes. The point of class separation is used as a global threshold with improved results.