A Segmentation Technique for Flexible Pipes in Deep Underwater Environments

This paper presents a novel segmentation technique for flexible pipes in deep underwater environments. It consists in finding regions of the pipe denominated vertebrae and has three stages. Firstly, the input image is pre-processed to reduce noise. Secondly, the pre-processed image is binarized by the proposed Multi-level Topological Binarization technique, resulting in an image which highlights (in white) the regions that are potential vertebrae (the blobs). Lastly, the pipe is segmented by finding the best sequence of blobs that fulfills a set of restrictions inherent to a pipe. For improving the technique robustness, an alternating pattern of white (a vertebra) and black regions is marked on the pipe. Results show that the proposed technique can segment pipes even under harsh conditions such as: low contrast between pipe and background; uneven illumination of the pipe; and high level of noise due to floating particles. In particular, the proposed binarization technique achieves valuable results without any parameter, whereas state-of-the-art techniques did not achieve the same quality even after their parameters were fine-tuned for each condition. Two blurring filters are used for pre-processing input image: a bilateral filter [5] removes small particles floating around, while preserving the vertebrae’s edges; and a Gaussian filter uniformly smooths the images. The proposed binarization technique finds peaks in the pre-processed image. This is done by "slicing" the image into a number of slices si,b (i is its index and b is its level). A tree of slices is then constructed by hierarchically connecting nodes of consecutive levels (Figure 1a). Intuitively speaking, peaks are prominences similar to the first one exhibited in Figure 1b. More formally, a peak is a sequence of nodes denoted by P = {nk}k=1 = n1,n2, . . . ,nm, such that: nm is a leaf node of the tree; nk, such that 1≤ k < m, is the parent node of nk+1; n1 is the only element that has a brother node, or n1 is the only element that has a parent with genus greater than zero, or n1 is the root node. Nodes n1 and nm are respectively named the base and the top of a peak. Given a tree of slices, one has to traverse it from leaves to root in order to find its peaks. For example, Figure 1a contains three peaks: the first one is P′ = s1,2,s1,3; the second is P′′ = s2,2,s2,3; and the third is P′′′ = s3,2,s3,3,s1,4.