A 2-d Filter Specification for Sonar Image Thresholding

We propose a new image sonar segmentation by combining two complementary competencies. On the one hand, following an image processing approach, we aim at partitioning raw image data to provide a binary image. On the other hand, we take advantage of technological knowledge such as the principle of sonar image formation. For sonar images, grey-level histogram generally presents a single mode which entails a poor separation of two theoretical modes related to reverberation and shadow subpopulations of the image. The separation of these two modes is of critical interest in a further description of objects from their cast shadows. In this paper, an optimal filter is specified by a criterion which aims at changing the statistical properties of each area while making threshold value selection from the histogram easier. While minimizing the output pixels variance, pixel values in each region concentrate around the respective average value while, simultaneously, two distinct modes appear on the histogram of the filtered image. The minimum value found between the two modes in the smoothed histogram leads to the searched threshold. In addition, we show how filter aspect depends as well on the image sonar resolution as on sonar parameters.