Optimization of a sonar image processing chain: a fuzzy rules based expert system approach

This paper presents an improvement of a sonar image processing chain. In order to extract significant areas, the first step of sonar image analysis consist of segmenting the image into three classes: shadow, echo, reverberation. Segmentation quality and precision are of first importance for performances of subsequent tasks. The original segmentation process was carried out by two thresholds that delimit boundaries between classes. To make the chain more robust, thresholdings are replaced by an unsupervised fuzzy clustering strategy. In spite of its efficiency, this method depends on the image histogram. To take into account possible changes in the image grey level distribution, the authors propose to adapt clustering by mean of learning rules. Last step is a post processing, called Nagao filtering, that improves the regions boundaries thanks to a context based segmentation.