(Introduction) Detection of defect indications in radiographic images is a topic of current research in non-destructive evaluation of materials and industrial testing. The limited signal-to-noise ratio of the radiographic images is due to the physical nature of the radiographic method as well as the small size of flaws and their poor orientation relatively to the size and thickness of the penetrated part. The basic approach to enhancement and binary segmentation of such flaw indications is the computation of difference images between the original image and the evaluated background image from the original one. Then, the resulting binary segmentation of flaw indications is made by a thresholding operation on the obtained difference image. In the framework of this approach, termed as a dynamic thresholding, the known methods differ by the manner of evaluation of the background intensity and selection of the threshold value for binarization [1-6]. For example, the simplest known method consists of a subtraction of the low-pass filtered image from the original image and comparing it with a constant threshold. It yields poor results on radiographic images due to the present noise and non-homogeneous background. Usually, the threshold value in these methods is determined from local or global histograms based on the assumption of a two-mode histogram [3-5]. However, the initial assumption of the two-mode image histograms in industrial radiographs is rarely fulfilled because of large intensity gradient values. Another and relatively new approach to flaw detection is the multi-resolution image analysis by using a wavelet representation [7-9]. This method has the advantage of being able to detect efficiently local objects, which are different in size and location. However, it still has the disadvantage of the previously mentioned methods of differentiation in case of low-contrast objects on a noisy background because it does not consider the shape constraints of local objects as well as an explicit model of the intensity function.
Relatively good results have been reported by using a model-based statistical approach to segmentation and detection of local objects in images with a robust statistical estimation of the underlying model parameters [9-16]. However, the statistical methods based on maximum likelihood decision making and hypothesis testing are not yet practical for this case study of radiographic images. Another concern of this approach is the computational expenses of the image analysis for detection of local objects because the statistical approach to detection is a computationally very extensive procedure. Adequate image modeling and estimationof model parameters play an important role in statistical approach to image filtering and binary segmentation. Several statistical image models have been proposed recently which are suitable for radiographic image modeling as well [17-19]. The method considered in this paper for detection of local objects based on local binary segmentation has several advantageous features. The model-based approach is used which exploits object multi-scale morphological representation in order to perform a time-efficient image analysis. The concept of a multi-scale relevance function has been introduced for efficient solution of the object detection problem, which allows a quick location of local objects of interest with different size. Local contextual information regarding already detected primitive patterns and shape constraints of the objects to be detected is taken into consideration during the final decision making.
FULL-TEXT - ABSTRACT