Multiscale Segmentation For SAR Image Based On Neural Networks

Synthetic Aperture Radar (SAR) has the attractive property of producing images in any weather condition and also in absence of sun light. So, SAR images are increasingly used in a wide variety of application fields ranging from surveillance, environment monitoring and civilian protection, indeed. For these applications, the segmentation can play a key role in the subsequent analysis for target detection, recognition and image compression. Because of the nature of the SAR instrument, SAR images contain speckle noise, complicating the segmentation of SAR images. By considering the statistical characteristic of SAR image and analyzing the merit and demerit of some kinds of algorithms, this paper proposes three new effective multiscale methods for SAR image segmentation based on combining multiscale method with neural networks.The details are as follows:1. Multiscale method for SAR image segmentation based on self-organizing feature map (SOM) neural network. Firstly, MAR (Multiscale Autoregressive) model is used for character extracting of SAR image, which are the input of neural network. Then, the structure and learning algorithm of SOM neural network are modified and trained so that it is qualified for SAR images segmentation.2. Multiscale method for SAR image segmentation based on probabilistic neural network (PNN). Firstly, MAR model is used for character extracting of SAR image too. Secondly, the training samples are clustered via K-means so that we can reduce the size of pattern layer and choose initial value of some parameters. Then, the PNN was trained by the algorithm we have proposed. Finally, the SAR images are segmented.3. Multiscale method for SAR image segmentation based on wavelet neural network (WNN). Firstly, wavelet is researched and morlet function was selected as the transfer function of hidden layer of neural network. At the same time, we research the initial value of parameters and learning algorithm of neural network. Finally, SAR images are segmented by the proposed method.4. Comparing and analyzing the three proposed methods from different aspects.In summary, this paper use multiscale method and neural network for the SAR image segmentation.The methods inherit the excellent strongpoint both of them and the experimental results show the methods are effective.