Automatic visual inspection capability has become an important key component in the automated manufacturing system. The most conventional sensor for visual inspection is CCD camera which provides an intensity/color image of a scene. One of the challenging tasks in visual inspection using CCD camera is to identify surface defects in an image with complex textured background. In microscopic view, the surface of real objects often shows regular or random textured pattern.
This paper presents a region identification method to extract surface defects in an image with textured background by using the space and frequency information at the same time. Dominant Frequency Map (DFM) is proposed to describe the frequency characteristics of every local region in an image coordinate. The method using DFM information is verified through a series of simulation and a real experiments. As a real application, defect inspection of a 14" TV CRT mold is performed. This method successfuly identifies infinitesimal surface defects, whose size is larger than 50㎛, of the mold. Also, a neural network algorithm is applied to the high-level recognition of the surface regions. The experimental results show that the DFM-based approach is less sensitive to the environmental changes, such as illumination and defocusing, than conventional vision techniques.