Image segmentation by nonsupervised neural networks
Jean-Claude Di Martino, Brigitte Colnet · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1994
In this paper, we propose a nonsupervised neural network approach to an image segmentation issue. The purpose is to extract spectral lines from sonar images. A Kohonen's self-organizing map is used to approximate the probability density function of the input data in a nonlinear way. The originality of our work with regards to the Kohonen approach is that constraints due to spectral lines features (temporal continuity, high mean energy), are encoded into the network. The process consists of two steps. First, searching for each point in the image whether a spectral line goes through this point. This step is achieved using a one-dimension map which self-organizes until a stable state is reached. The second step consists in evaluating whether network topology recovers spectral lines properties. For this purpose, we define an objective function which depends on neurons mean energy and global curvature of the network seen as a topological set of units. This process enhances spectral lines perception in a noisy image and has been successfully applied to a set of lofargrams with different signal to noise ratio.