DEVELOPMENT OF A LOW-COST VISION SYSTEM FOR FINDING CONTOUR AND SURFACE DEFECTS ON CAST IRON ENGINE COMPONENTS

Abstract: The thesis aims to develop an image processing algorithm for a 2D vision system for quality inspection of cast iron components (size 250x190x120mm). The system is designed to be low cost and be an easily implementable solution for an existing production line to complement or replace part of manual quality inspection. The defects are mainly caused by gas porosity and molding errors and can have a very different nature starting from small cavities and chipped edges ending with large portions of missing material. Thus, the methods used in the algorithm must guarantee successful detection of very different types of defects while handling natural shape deviations in contours. Several known image processing methods were tested and further developed to fit this specific application. First, suitable candidate methods were selected, implemented in MATLAB® and tested on a small set of manually taken test images. The best performing methods were then further developed into a fully functional beta algorithm that was then tested on real production line using prototype hardware. To understand the system’s capabilities a blind experiment with quality inspectors was carried out. The algorithm uses Canny edge detection operator for obtaining the contours together with several stages of dynamic masking and morphological processing to eliminate false edges and non-important contours. Intensity mapping and filtering is then applied to eliminate traces of cutting fluid on the part as well as occasional high-contrast milling marks. Surface defects are then measured and compared with an inspection standard to decide if part is approved or not. Thereafter the contours of the part are checked for deviations caused by porosity and damaged molds. Shape signatures and segment-wise alignment with a template signature is the method that proved to be sufficient in most cases. The method has shortcomings in terms of determining the size of deviations in certain segments due to the complex shape and significant natural variations in contours. Ideas are proposed for improvement of the system’s accuracy and robustness.