Sonar image recognition using synthetic discriminant functions implemented with the Karhunen Loeve transform

This paper discusses modified synthetic discriminant functions (SDF) using a Karhunen Loeve transform (KLT) used for improved sonar image recognition. The SDF filter synthesis involves using the whole image which in turn creates redundancies in the distinguishing features. A number of different schemes have been used to try to reduce the data in SDF filters in order to make them more practical, efficient, and reliable. The KLT is one method to reduce the redundancies in a set of training images to create a new data matrix. This data matrix has a new coordinate system in which the axes of the system are in the direction of the eigenvectors of the covariance matrix of the training set. With the realigned data found in the data matrix, the principle component images can be extracted. Principle component images are comprised of the variations of the original training images. This minimizes the training data to the necessary information that is needed for image recognition. The principle component images then become the training set to be used in the SDF filter. Because the KLT allows for the reduction in redundancies by examining only the variations of this new training set, it will increase the correlation found in this implementation of the SDF filter.