IMPROVED APPROACH FOR INVARIANT AIRCRAFT TYPE RECOGNITION

Abstract

Despite a great deal of efforts to automate the aircraft recognition process aircraft recognition remains a challenging problem. The majority of the aircraft recognition methods assume the successful isolation of the aircraft portrait from the background, and only a few have actually addressed real world concerns, such as clutter and shadows. In this thesis, I present an automatic aircraft recognition system, which shows improved performance because of ring projection and dual tree complex wavelet. This system assumes from the start that the image could possibly be degraded, contain clutter, shadows and blurring. Feature extraction is a crucial step in invariant pattern recognition. Among all existing feature extraction techniques, ring-projection has been selected for invariant pattern recognition in [28]. This is because it is invariant to translation and scale of the patterns. In addition, the ring-projection transforms the feature space from 2-D to 1-D, which reduces the processing time substantially. This makes it suitable practical approach in real-time applications. The dual-tree complex wavelet transform is applied to the ring-projection signal in order to extract shift invariant features at different resolution scales. The reason why we choose the dual-tree complex wavelet transform is because it has the approximate shift-invariant property, which is very important for pattern recognition. Figures 4-7 show the correct recognition rates of the descriptor and the Fourier transform for different rotation angles at SNR=10, 5, 3, 1, respectively. From these figures, it can be seen that this descriptor is much better than the Fourier transform especially for high noise levels. If more representative templates (not only 4 templates) are used to represent an aircraft type, the three failure cases will be tackled well.