As the name of the paper suggests, the work described here is a study of the use of features derived from Gray Level Cooccurrence (GLC) matrices to detect spatially compact objects in images. Coocurrence matrices are commonly used in texture classification. As the authors point out, however, GLC matrices contain information about both the edges and the homogeneous regions in the image; hence, they should be useful in object detection.
This paper follows a standard methodology in pattern recognition to determine the classification power of a set of features. The paper could easily be used as a paradigm to define this methodology. First, the features one wishes to study are defined. Here, a number of features are defined based on properties of the GLC matrices of sample image regions. Next, the training and testing data are specified. For their data, the authors chose to use regions from visual images of aircraft. Then the form of classifier is specified. A supervised parametric classifier is used: the features computed for object regions are assumed to come from a Gaussian (normal) distribution with unknown mean and covariance matrices. A Chi-squared test is used to determine if an unknown sample region can be rejected as having features compatible with the estimated normal distributions. Finally, the feature selection process is defined. The authors use a sequential jackknife (leave-one-out) procedure, which estimates the probability of error of the classifier when using a continually growing subset of the features.
The features the authors compute from the GLC matrices appear to work reasonably well for their aircraft detection problem. It would have been more informative, however, to see the results compared to other standard detection techniques. Not surprisingly, since GLC matrices are defined based on an orientation parameter, the orientation of the object to be detected is critical. The authors suggest using a series of decision modules, each “tuned” to objects of the correct orientation.