An Improved Graph-Based Image Segmentation Algorithm and Its GPU Acceleration

The paper proposes an improved graph-based image segmentation algorithm. We introduce a new merging mechanism based on minimum region size. The mechanism not only limits the minimum region size but also avoids very large regions which may cause over-merging. We can obtain better segmentation results utilizing our method which highlights the characters of single object and is more applicable for image feature extraction. In addition, we accelerate the algorithm on GPU (Graphics Processing Unit), which achieves large speedup (about 3.52 times) and comparable performance. Moreover, the advantage in speed is more obvious with the image resolution increasing. It can meet the real-time requirements to a degree.