Performance analysis of image thresholding: Otsu technique

Abstract Image thresholding is usually applied as an initial step in many algorithms for image analysis, object representation and visualization. Although many image thresholding techniques were proposed in the literature and their usage is well understood, their performance analyses are relatively limited. We critically analysed the feasibility of successful image thresholding under a variation of all scene parameters. The focus is based on Otsu method image thresholding technique since it is widely used in many computer vision applications. Our analysis based on Monte Carlo statistical method shows that the success of image segmentation depends on object-background intensity difference, object size and noise measurement, however is unaffected by location of the object on that image. We have also proposed a set of conditions to guarantee a successful image segmentation. Experiment using real-image data was set up to verify the validity of those conditions. The result demonstrates the capability of the proposed conditions to correctly predict the outcome of image thresholding using Otsu technique. In practice, the success of image thresholding could be predicted beforehand with the aid of obtainable scene parameters.