Application of time-frequency techniques to detection of minelike objects in side scan sonar images

Abstract

Wavelet analysis techniques were applied to side scan sonar images for the detection of minelike objects. A subwindow of sample images were scanned and compressed using a best level orthogonal wavelet basis. The compressed images were then projected onto a dictionary of target and nontarget examples in wavelet space, using the matching pursuit algorithm of Mallat and Zhang to find the most likely match. With the best level basis, however, signal representations are dependent on their location in space. Because of this, shifted versions of the target examples were included in the dictionary to improve the likelihood of a good numerical match. Shift-invariant wavelet techniques, which provide for target translation while maintaining the same numerical descriptors, were also studied with classifier performance and computational complexity being evaluated. This allowed best bases representations to be applied in the classification algorithm. The overall detection performance was further refined by use of a neural network operating on the set of possible mine detections, which had originated from the matching pursuit classifier.