Abstract

Active IR polarimetric sensing has been successfully applied for the remote sensing of man made objects and, particulary, of buried mines However, the scattering and power/SNR constraints require near overhead viewing. In vontrast, passive polarimetric sensing allows detection with much more operationally convenient arrangements which is highly desirable when working in mined lands This kind of sensors detect two types of polarized light: thermal self-emission and reflected solar energy [1]. Thus, information about temperature differences due to the different thermodynamic properties of the buried mine and the surrounding background can be extracted. The thermal contrast of an object with respect to its environment is a function of both its temperature and of the emittance of its radiating surface, and varies in proportion to the duration of heating In this work, a new approach for detecting buried antipersonnel mines based on the dynamic behaviour difference is presented. The basic idea consists of using a sequence of images of the same piece of land at different time intervals which are applied as the input of a reconfigurable cellular neural network (CNN) architecture. Then, a learning algorithm is applied, namely Genetic Algorithms (GA), that optimizes both the network parameters and the network topology that best fit the desired behaviour [2]. Obviously, in order to do this, during the training phase of the network an input sequence of images together with the corresponding desired output must be supplied This implies the need of an a priori knowledge of the mine position for the training sequence After the training has finished, the system can be used with arbitrary sequences where the exact mine location remains unknown The basic operation of the system is schematically shown in the figure Here a sequence of four images is used that act as external inputs of the system Then, GA is applied to reconfigure the network structure in order to find the topology that best fits the problem under consideration. This process of reconfiguration consists of selecting the adequate number of layers and connections among them together with the network weights that describe the operation of the system. The output of the system will be the extracted mine (black pixels) over a white background. Images were supplied by the ETRO-IRIS group of the Department of Electronics of the Free University of Brussels (http://etro.vub.ac.be/minedet/)

© 2000 IEEE