Improvement of X-ray castings inspection reliability by using Dempster–Shafer data fusion theory

The aim of this work is to improve the classification of defects in X-ray inspection by developing a new method based on Dempster–Shafer data fusion theory where measured features on the detected objects are considered as information sources. From the histogram of features values on a learning database of manually classified objects, an automatic procedure is proposed to define a set of mass functions for each feature. The spatial repartition of features is divided into regions of confidence with corresponding mass functions. A smooth transition between regions is ensured by using fuzzy membership functions. The whole process is carried out without any expert intervention. Validation takes place on a testing database. Data fusion leads to a significant improvement of classification performances with respect to the actual system.

Research highlights
► Data fusion improves defect classification performance.
► Features are considered as sources of information in a data fusion framework.
► Mass functions are computed automatically from the features values histogram.