Simple Linear Iterative Clustering (SLIC) and Graph Theory-Based Image Segmentation

Introduction

Image segmentation is one of the most important concepts in image processing and computer vision (Chowdhary & Acharjya, 2020; Chowdhary, 2011). In the image processing approaches, regions with high-frequency texture will be estimated by stipples, whereas regions with lower-frequency texture will be estimated by closed shapes (Chowdhary, Goyal & Vasnani, 2019). Creation of the closed shapes and taking a decision about the separation of an image into regions depend on covering pixels with similar properties. Therefore, there is a requirement for a proper image segmentation algorithm. Simple Linear Iterative Clustering (SLIC) is one such segmentation algorithm which is suitable for splitting the image into proper regions. SLIC is efficient and this produces regions that adhere well to edges in the image. The pixels in SLIC clusters are combined in five-dimensional colour and they are plane spaced to efficiently generate compact and uniform superpixels (Stutz, Hermans & Leibe, 2018). The computer vision applications rely increasingly on superpixels but constitution of a successful superpixel algorithm does not always straightforward (Achanta et al., 2012). It is possible to treat a superpixel as a group of pixels that are identical in location, color, texture, etc. Superpixels can engage image redundancy and they transform computation at the pixel level into an operation at the region level, which can greatly reduce the complexity of subsequent tasks of image processing. In different image processing applications, such as image segmentation, saliency detection and classification, superpixel segmentation has become a significant pre-processing stage (Wang, Peng, Xiao & Liu, 2017).

It is important to describe current superpixel segmentation methods into three major categories: the spectral-graph-based method, the gradient-ascent-based method and the optimization-theory-based method. The most commonly used superpixel form is SLIC. Generally, it is a technique of local k-means clustering (Wang, Peng, Xiao & Liu, 2017; Chowdhary & Acharjya, 2016; Das & Chowdhary, 2017; Chowdhary & Mouli, 2012; Chowdhary & Mouli, 2013; Chowdhary et al., 2020). Galasso, Cipolla and Schiele (2012) show that frame-based superpixel segmentation combined with a few motion and appearance-based affinities are sufficient to obtain good video segmentation performance. This improves the performance for video sequences due to motion-clues. An image segmentation benchmark to videos allows coarse-to-fine video segmentations and multiple human annotations (Galasso, Cipolla & Schiele; 2012). The requirements for superpixels are mentioned in Table 1.