The concept of super pixels at the earliest in 2003 by Ren et al, the super pixel algorithm to extract image perception in meaningful regions, can be used to replace the pixel grid rigid structure, using super pixels instead of pixel operation, can speed up the existing pixel based algorithm to improve some of the results for the past 10 years.
Study on super pixel has been developing rapidly at home and abroad, has become a key technology in the field of computer vision research for many applications in image segmentation, many existing segmentation algorithm based on graph theory, such as Ncut, its cost will increase as nodes in the graph and become expensive, because this algorithm will limit the size of the image size for some specific applications, such as electron micrographs of mitochondria segmentation, the image size is large, the image segmentation based on pixel grid is very difficult, the image pixels around the super pixel aggregation using SLIC algorithm, then each pixel as each node in the graph to achieve image segmentation, it can effectively reduce the complexity of the image, the segmentation becomes easy to handle.
As shown in Figure 11, using the SLIC supervoxels method, large size 3D image segmentation to billions of pixels, and the algorithm is of low complexity, reduce the memory requirements, can significantly improve the performance of Kohli et al. To solve the problem of how to will belong to the same label segmentation fragment with the same object problem.
In the human pose estimation, Mori et al. First the image is divided into blocks of super pixels or larger, detect and locate the contour of the human body joints and limbs, and then all parts of the body are combined with Mori super pixel segmentation pretreatment, improves the search pattern in the image of the efficiency and accuracy, and in static images in the human pose estimation is achieved good results, as shown in figure 12.
In the field of target tracking, Wang et al. Proposed an object tracking algorithm using hyper pixel to extract object structure information from the perspective of intermediate level vision.
They use a super pixel based appearance discrimination model to make the tracker distinguish the target and the background through the intermediate level line cable. Then, the tracking task turns into calculating a target background trust value, and the best candidate result is obtained by the maximum a posteriori estimation.
The tracking algorithm can effectively deal with deformation, target tracking in occlusion and occlusion, as shown in Figure 13 Zhou et al. Proposed super pixel driving level set tracking algorithm, the definition of a speed function to capture the correlation between super pixel or target and background, the algorithm has good robustness and high efficiency of Liu who realizes tracking of multiple vehicles in the real world in traffic video, the semantic information is introduced into the super pixels, effectively solve the occlusion problem and frequent cross of different vehicles.
Wang et al. Explored the target tracking problem by exploring the super pixels based on the visual information around the target, and proposed an appearance model composed of several components
The algorithm is better than other algorithms in the case of object deformation and occlusion
Can also be used in other aspects of the super pixel image processing task, Gu et al. Super pixels used in image scene classification, the image segmentation for super pixel blocks, and image SIFT feature extraction, the formation of contextual visual descriptor, and then use the space in Pyramid to represent the image and classification method.
Tighe et al. Applied super pixel to scene component analysis, and proposed a simple, non parametric and efficient image analysis method. Fulkerson et al. Expounded the method of locating target and segmenting target class by using super pixel in image. Their experimental results on Graz-02 and PASCAL VOC 2007 datasets are much better than many existing image segmentation methods.