A RBFN Perceptive Model for Image Thresholding

Abstract

The digital image segmentation challenge has demanded the development of a plethora of methods and approaches. A quite simple approach, the thresholding, has still been intensively applied mainly for real-time vision applications. However, the threshold criteria often depend on entropic or statistical image features. This work searches a relationship between these features and subjective human threshold decisions. Then, an image thresholding model based on these subjective decisions and global statistical features was developed by training a Radial Basis Functions Network (RBFN). This work also compares the automatic thresholding methods to the human responses. Furthermore, the RBFN-modeled answers were compared to the automatic thresholding. The results show that entropic-based method was closer to RBFN-modeled thresholding than variance-based method. It was also found that another automatic method which combines global and local criteria presented higher correlation with human responses.