Cluster

Project summary:

Superpixel is a crucial image preprocessing technique that receives increasing attention in many multimedia and computer vision fields, such as salient object detection, image segmentation, and visual tracking. The technique targets at oversegments an image into a number of perceptually meaningful regions. Compared with the rigid pixel representation of images, superpixel representation contains less redundancy and agrees well with human vision. As a preprocessing step, superpixel can provide a substantial speedup for the subsequent operation, since it greatly reduces the number of elements to be processed.We deals with the superpixel segmentation problem using a powerful global optimization technique: Differential Evolution (DES). The algorithm mimics the process of nature evolution to realize efficient optimization, and it poses no restrictions on the form of objective functions. This way, we develop a novel and comprehensive objective function considering both local and global costs in the segmentation, including within-superpixel error, boundary gradient, and a regularization term. The proposed method can produce superpixels in a computational time linear to the image size. Experimental results on the BSDS500 dataset validate the competitive performance of our algorithm in terms of boundary adherence and segmentation capability.

Research highlight:

DES optimizes the global optima in a more direct way by means of global optimization. We design a Boundary Gradient term to evaluate the adherence of superpixel boundaries to the high gradient components in the image. Further, to enforcing generating homogenous superpixels, we introduce a Regularizer to measure the global variance of superpixel sizes. The two global terms, as well as the traditionally used local cost named Within-Superpixel Error, is considered in the objective function. The optimization is then accomplished by Differential Evolution (DE), a powerful stochastic global optimization algorithm mimicking the nature evolution process. Owing to the low complexity of DE, the proposed algorithm satisfies the computational restriction that it can produce promising superpixels with a linear computational complexity.

Figure 2. Performance comparison curves of different superpixel algorithms. (a) Boundary Recall. (b) Undersegmentation Error. (c) Executing Time per Image.

Figure 3. Achievable segmentation results based on DES, EOpt0, SLIC, and TP.

Project summary:

Gong Y J, Zhou Y, Zhang X. A superpixel segmentation algorithm based on differential evolution[C]//Multimedia and Expo (ICME), 2016 IEEE International Conference on. IEEE, 2016: 1-6.