A robust texture segmentation algorithm based on the Fourier spectrum analysis

摘要

This paper presents a segmentation algorithm for image processing, which is able to distinguish the background's area from the object's area based only on their texture characteristics. The main goal is to develop a segmentation algorithm which does not need any prior information about the object (form, colours, textures, brightness etc.) and which is also robust against industrial conditions (such as shadow effects and electromagnetic noise). The developed solution is based on a predefined and structured background (with a specific frequency in the grey values). The background frequency is compared with the object's texture frequency using a Fourier analysis together with a classification algorithm. Two different types of classification algorithms have been applied: statistical covariance analysis and Neural Networks. The algorithm has been evaluated using real textile images under distinct conditions showing a very appropriated result with a segmentation error smaller than 0,5% in the average. The paper finishes with a set of conclusions and perspectives for future works.