Segmenting images with support vector machines

The aim of this work is to propose an original image segmentation methodology to detect and localise objects or patterns in an image. This new technology has two parts: (a) the main module is a SVM neural network whose goal is the image segmentation in order to detect and localise objects having regular patterns (represented by a block of pixels), and then, (b) a simple morphological processing, to eliminate isolated misclassified pixels. The importance of this methodology is highlighted with the results obtained in the recognition of 2D symbolic codes. Another advantage of our algorithm is its regularity that may be exploited to propose a parallel hardware architecture.