A Multi-View Sonar Image Fusion Method Based On The Morphological Wavelet And Directional Filters
Zhang Zhigang,Bian Hongyu,Song Ziqi
DOI: https://doi.org/10.1109/coa.2016.7535779
2016-01-01
Abstract:In underwater object detection and identification of an unmanned underwater vehicle (UUV), high-frequency 2-D imaging sonar is an effective device for measuring distance accurately. Forward-looking sonar records a sequence of images of the object's surfaces, which continuously change as the UUV moves around the target. A single image usually describes partial structures or local regions of a large object surface, such as for ship hull inspection, and the surface detection of a pier or dam. To improve the capability of object detection and identification, we focus on the details and textures of sonar images, and use the image fusion method to take advantage of complementary information among redundant images. In our work, we consider the features of a sequence of images taken from different views, and which have different target intensities, object shapes and noise distribution. We present a sonar image fusion method based on directional filters banks and morphological wavelets, which combines the features of multi resolution wavelet analysis and nonlinear filters. Firstly, the noisy sonar images are transformed into the morphological wavelet domain, which can effectively decrease image noise. Then the high frequencies of the source images are fed into a directional filters bank, which uses the directional decomposition approach to provide exact details of the image whilst retaining the unchanged low frequency content. Finally, we fuse the multiscale and directions parts into the transforming domain, and reconstruct the fusion image. The result of applying this method to a sequence of sonar images from multi-views indicates that the fusion image can effectively describe the extra details in source images in terms of multiresolution and direction. It is also good at suppressing noise, especially for images with a higher noise level. Comparison of the experiments and real data supports our conclusions from subjective and objective evaluation, and shows that the regions in the fused images are effectively repaired and image quality is evidently improved.