Path Reducing Watershed for the GPU

Abstract

The watershed transform is a popular image segmentation procedure from mathematical morphology used in many applications of computer vision. This paper proposes a novel parallel watershed procedure designed for GPU implementation. Our algorithm constructs paths of steepest descent and reduces these paths into direct pointers to catchment basin minima in logarithmic time, also crucially incorporating successful resolution of plateaux. Three implementation variants and their parameters are analysed through experiments on 2D and 3D images; a comparison against the state-of-the-art shows a runtime improvement of around 30%. For 3D images of 128 megavoxels execution times of approximately 1.5-2 seconds are achieved.

Facial Image Segmentation by Integration of Level Set and Neural Network Optimization with Hybrid Filter Pre-processing Model

“…The performance analysis of our proposed work is conducted with some of the performance metrics such as sensitivity, specificity, accuracy, Dice coefficient, [40,41] Jaccard coefficient [42,43] and compared with some of the prior works such as FCN, [44] GF, [45] MSFS, [46] GA. [47] The following section of this article encloses the description of performance metrics and their comparative analyses.…”

Parallel Partitioning: Path Reducing and Union–Find Based Watershed for the GPU

“…The computational bottleneck of this algorithm is step S2 which, in the worst case, requires a number of iterations linear in the image size. We have shown [9] three implementation variants for this step: synchronous, block-asynchronous and balanced. The synchronous variant is the slowest due to all CUDA threads synchronising after every iteration.…”

“…We converted the BSDS500 images to greyscale without any further processing before running the watersheds. In addition, we include three 3D images (128 megavoxels each) of resolutions 320×320×1250, 500×500×512, 4000×4000×8, and a single 3D image (800 megavoxels) of resolution 4000×4000×50, all constructed from the same microCT image [9]. The 4000×4000×8 image is also interpreted, in separate experiments, as a set of eight 2D images.…”

“…Reviews of definitions and sequential algorithms can be found in [6,7,8]. We have explained elsewhere [9,10] the many and varied uses of this class of method. Suffice to say, for brevity, that these are used for partitioning images into regions with pixel similarity, as seen in Fig.…”

“…With the fast development of general-purpose computing on GPUs, a number of parallel watershed algorithms have been introduced for that setup [11,12,13,14,15,9]. In our previous work [9] we presented experiments where our PRW algorithm [16] outperformed the DW algorithm [12,17]-the state-of-the-art at the time. We propose two novel parallel watershed algorithms designed for GPU implementation.…”