Super Pixel Segmentation Method Based On Gradient Descent

Super Pixel Segmentation Method Based On Gradient Descent

Nov 17, 2017

1.1  Watershed method

1.2  Based on Mean- shift method

1.3  Turbopixels method

1.4  SLIC method

2.  Comparison of experimental results

  In order to further understand the performance of several super pixel segmentation methods, this paper conducted a comparative experiment in Berkeley benchmark standard data sets, the verification algorithm including entropy rate (ER) Ncut- based (Ncut- B) su-perpixel lattice (SL) Turbopixels (TP) and SLIC of each picture size is 321x481, divided into about 200 a super pixel results,

  As shown in Figure 10 shows that ER algorithm can keep the image edge information, but the super pixel shape irregular, each image segmentation time is about 1.16sNcut- B algorithm can effectively maintain the image boundary, and with the increase of the number of super pixels, the super pixel shape will be more regular, but the processing time of the image segmentation is longer, the size of the 321x481 image to 200 a processing speed of super pixels around 2min SL algorithm is faster than that of about 0.36s can be divided into image grid, but the quality of the segmentation results it severely influence the input boundary map of TP and SLIC algorithm can produce regular and compact super pixels but, the method of TP boundary remain poor,


 SLIC from the boundary to maintain speed and super pixel shape evaluation is higher, compared with the expected segmentation result in memory 3.00GB processor Intel Core 2 machine, the time comparison of these various super pixel segmentation algorithm as shown in table 1.