A Real-Time Superpixel Algorithm based on Two-Layered Grid Seams

• Main Authors: Jia-Le Lin, 林家樂
• Other Authors: Chang Hong Lin
• Format: Others
• Language: en_US
• Published: 2015
• Online Access: http://ndltd.ncl.edu.tw/handle/49660710645165482110

碩士 === 國立臺灣科技大學 === 電子工程系 === 103 === In recent years, thanks to the great growth of the PC and smart mobile device development, real-time image processing applications are also widely discussed and proposed. With the advanced camera, the resolution of the image may achieve Full HD (1920×1080 pixels) or even higher easily. Consequently, the image can contain more information, which means that we can retrieve more precise features from the image, but the computational load is proportional to the image resolution most of the time. Superpixel algorithm integrates the adjacent pixels which have similar color to superpixels. It resembles that superpixel algorithm reduces the resolution of image but still maintains the image structures. Therefore, superpixel algorithms also are widely discussed and many superpixel methods are proposed. In this thesis, we propose a fast superpixel method which has high accuracy. The vertical and horizontal First-Seams produces roughly grid-shape image structures. Then, the vertical and horizontal Second-Seams can figure out the relatively delicate structures. The Dynamic programing is used to find the optimal seam. Hence, the proposed method can achieve a high accuracy, and it only need about 7 milliseconds to complete the computing when it is processed on an Intel i7-4770 CPU platform with a single thread. By the comparison with other methods, it shows that the proposed method is faster and more accurate than most of other methods. Finally, we analyze the factors which boost the accuracy and sum up the ways which may improve the superpixel method and should be the interesting topics for future work.