Unsupervised registration of textured images: applications to side-scan sonar

Abstract

Sonar images are highly textured images and therefore mislead most of the classical registration algorithms. Registration is a critical step for the creation of high-resolution accurate mosaic images of the seafloor required for seabed analysis and classification. In the past concurrent mapping and localisation have successfully been used but the detection and association of landmarks have been proved difficult and been done manually. However, such methods are time consuming and lack robustness. Landmarks are not regularly present in the images and their localisation is prone to errors. As a consequence, global methods using whole images are preferable. These methods were extensively studied in the recent years and successfully applied to multimodal medical image registration. Unfortunately, the similarity metric between images they rely upon cannot cope with highly textured images. To overcome this issue, textural features must be extracted to highlight similar regions of the images. Registration of these feature maps works but remains sensible to the feature selection and their relation from one modality to the other. An alternative approach is proposed in this paper. Mutual information is calculated from all the features and global registration can be achieved directly. Solely an approximation of MI can be obtained but the performance of this algorithm are equivalent to exact approach and robust to feature selection. This method has been successfully applied to textured images (side-scan sonar) but is also applicable to multimodal images such as bathymetric and sonar data.