A new method, based on an Infomax Learning Algorithm and neural networks, for the pattern extraction in Synthetic Aperture Radar (SAR) images was developed. SAR Images can be obtained from the radar backscatter that is fully dependent on the surface conditions of the target. Then, it can distinguish the target from non-target by extracting the difference in patterns generated by the backscatters.
The difficulty is that the difference of geometric pattern's characteristics of these targets is unknown beforehand, and then it is not easy to adapt simple filtering to divide these patterns.
We tried to develop the computer algorithm simulating the pattern extraction system in reference to the excellence of human vision's recognition ability. We propose the classification method based on image patterns using Infomax Learning Algorithm which simulates human visual cortical neurons as the possibility to automatically extract the specific area from an image. This algorithm considers a neural network model of visual area for modeling human visual recognition process, learns neural networks based on the idea of Infomax (Information Maximization), and learns local image pattern to obtain the weight patterns required for classification, which are approximated by Gabor function. We attempted to detect airplanes and aprons in SAR image taken at Kansai International Airport by the proposed method. As the result of adopting Infomax for the SAR image based on the improved learning method, we successfully demonstrated to detect airplanes from other various, complex artificial constructions, which are specific to the airport, in the image. Hence, it was demonstrated that target was automatically detected by learning of its patterns without any mathematical definition of equations for the pattern.