Combining the Delaunay Triangular Mesh for Image Segmentation

To reduce noises during image processing, a Delaunay triangular mesh combined with image segmentation algorithm is presented. Firstly, the grayscales of pixels can be viewed as the elevations defined the lattices of the pixels, then a Delaunay triangular mesh is built on the three dimensional of the given image. The projection of the Delaunay triangular mesh partitions the domain of the image into sub-regions which are coming into superpixels. The grayscales of pixels in each sub-region is respectively averaged to generate an averaged image. Finally, segmentation is carried out on the averaged image rather than the original image. Take the classical FCM (Fuzzy C-means) which is based on the platform of Intel(R) Core(TM) 3.20GHz/2G of memory/Matlab2015a as an example, the image segmentation experiments are done with Berkeley benchmark images and remote sensing images. Generate a simulated image with three homogeneous regions and adopt the contrast methods as well as the proposed method to do the segmentation experiments. The user accuracy、product accuracy、overall accuracy and Kappa coefficient of the three homogeneous regions (1-3) of the contrast method 1（ISODATA algorithm based on ENVI platform）are 74.87%/55.72%/73.64%、83.98%/37.87%/85.38%、70.25%和0.54; the user accuracy、product accuracy、overall accuracy and Kappa coefficient of the three homogeneous regions (1-3) of the contrast method 2（FCM algorithm based on median filtering）are 98.16%/70.54%/99.73%、88.10%/99.16%/87.87%、89.93%和0.84; the user accuracy、product accuracy、overall accuracy and Kappa coefficient of the three homogeneous regions (1-3) of the proposed method are 96.25%/80.35%/99.49%、99.76%/94.53%/86.67%、93.30% and 0.89. The results of quantitative and qualitative experiments suggest that the validity, reliability and accuracy of the proposed algorithm.