The existing external inspection and X-ray transmission inspection methods can not produce some reliable results for the case of the internal shape of the 3-dimensional object. If the cross-sectional image for the plane-of-interest of an object can be obtained, the cross-sectional shapes of the object are recognized from this image.
This thesis is a study on the technology of realizing arbitrary cross-sectional images in a digital tomosynthesis system by image reconstruction using an image set of the focal plane of an object. First, a new digital tomosynthesis system of the object-detector synchronous rotation(ODSR) type was designed for acquiring the digital X-ray images. For improving the quality of the images acquired by this system, the image distortion, the intensity attenuation, the sharpness of the cross-sectional image and the artifact phenomenon were mathematically modeled and the improvement algorithms were suggested. And the evaluation index to estimate quantitatively the image quality was defined, and the improvement results by the simulations were shown in comparison with the experiments.
The mathematical models for image reconstruction to realize the cross-sectional images of arbitrary planes of 3-dimensional objects using the image set for a focal plane were derived. Using these models, the realized results of the cross-sectional images for arbitrary angle planes and height planes were represented by the computer simulations. Also the experimental images were acquired by the ODSR system designed in this work for the same objects as used in the simulations, the cross-sectional images of arbitrary planes by image reconstruction from the image set were realized and compared with the simulation results.
As the practical application, the study on the defect inspection of the soldering joints of ball grid arrays was performed. The image set of the horizontal focal plane of the joint was acquired, the cross-sectional images of arbitrary planes were realized by image reconstruction using this image set. The 7 features were extracted from the intensity profiles of the cross-sectional images. A number of feature sets extracted from the experimental data of 540 sets were classified to 4 different classes by the semi-linear feedfoward network classifier and the percentage of the correct classification was shown about 96.1%.
