This paper presents a defect detection method based on two dimensional Fourier transform. In automatic visual inspection in factories, the displacement between two inspected images contains both parallel shift and rotation, which causes the inspection process to be highly time-consuming. In the proposed method, the subtraction between Fourier transform magnitudes of two input images is calculated for detecting differences between the corresponding pixels. Due to the shift- invariant nature of magnitudes in Fourier transform spectrum, the method does not require an image registration process for parallel shift. For images rotated by a small angle, a correlation based on one dimensional Fourier transform is calculated to estimate the angle.
Detection experiments are executed on Mac Pro 2.66GHz using images with 256 × 256 to 4000 × 2672 pixels. For the largest resolution images, the method detects differences in 2.39 seconds. The experimental results show that the proposed method has sufficient detection precision and faster computation speed than existing visual inspection methods. Estimation of specular reflection on PCB surfaces and further acceleration of the calculation are to be implemented.
View full abstract