Long-Range Target Classification in a Cluttered Environment Using Multi-Sensor Image Sequences

Accurate identification of unknown contacts is crucial in military intelligence. Automated systems which quickly and accurately determine the identity of a contact could be a benefit in backing up electronic signal identification methods such as Identification Friend and Foe (IFF) systems. Radio Detection and Ranging (RADAR) images are often undesirable in military applications since they reveal the location of the imaging system. So we explore the use of visible and infrared images of which are generally more consistent than RADAR images and for which it is easy to compensate for environmental effects. Recent advances in visible and IR imaging technology improve the ability to observe objects at very long distances, but it is still militarily desirable to stay away as far as possible from potential enemy ships that may be observed at different viewing angles. A Principal Component Analysis (PCA) based pattern recognition technique is presented in this thesis for small boat classification from visible and IR images. Extracting features from naval ship images is a difficult task to solve since the background may contain rough-textured regions. Conventional edge and comer operators tend to fail due to high contrast in these regions. Appropriate preprocessing steps are necessary to separate the object from the background. Subsequent feature extraction techniques can be applied to the whole object region. Two efficient segmentation algorithms are used for the extraction of small boat regions from visible and IR images. A popular and convenient approach named Adaptive Progressive Thresholding (APT) technique is used for segmentation of boat region from visible images. Infrared images in general are not sharp in displaying objects compared to the visible images. In fact, the noise components present in these images are always much greater than typical visible images. Therefore, pure edge-based image segmentation will not be sufficient in this case. We applied SUSAN-APT and Graph cut image segmentation algorithms for small boat region extraction from IR images. We obtained 86% accuracy in classification rate in visible images and 77% in IR images. The classification accuracy can be improved by using more training images captured in every 15 degrees viewing angle for all classes as opposed to only 5 training images used in this work for every class. This research can also be easily applied to air and land based target classification systems.