Quantitative Analysis of Image Segmentation Algorithms: A Statistical Perspective

Abstract

These Image segmentation process includes image pre-processing, background removal, foreground detection, outlier removal and post processing. In order to develop an efficient image segmentation system, it is mandatory to design these sub-processes with utmost efficiency. Image segmentation systems can be designed to be application specific or application independent depending upon the design of these internal modules. Algorithms like kMeans, fuzzy CMeans, etc. are generic, but the output of these algorithms must be tuned depending upon the given application for efficient segmentation. For instance, in order to effectively segment leaf imagery, the output clusters of kMeans must be checked for green coloured values, and clusters where green colour is prominent must be extracted for segmentation. The same task can be done via the use of Saliency maps, Grey level co-occurrence integrated algorithm, etc. by tuning their internal parameters. Thus, there are a wide variety of similar algorithms developed In order to reduce this ambiguity, this text reviews different image segmentation algorithms, and compares their statistical performance in terms of peak-signal-to-noise-ratio (PSNR), delay needed for computation, minimum mean squared error (MMSE), most probable application, etc. Moreover, this text also evaluates certain nuances, advantages and drawbacks of these algorithms, which will assist researchers to select the best algorithm set based on their application. This text also recommends certain improvements which can be done in these algorithms, in order to improve their performance via fusion, cascading and ensembling.