Tape substrate pattern of ultra-fine pitch circuit less than 10 micrometers in pattern width, is required to be inspected through high resolution optics. In the process of picking out defects at the level of the critical dimension through image processing, however, trivial blemishes formed by dust or micro particles may be detected simultaneously. This leads to unnecessary work on the part of operators reviewing and verifying the additional detected points. To maximize the efficiency of the inspection process, we need to identify and classify the defect candidates whether it is a real pattern defect or simply a trivial blemish by dust. Since a real defect arising from under or over etching bears inherent features in shape and brightness, it can thus be discriminated from other trivial blemishes. In this article, we propose an image feature based defect classification method, where proper measures were obtained from a series of image analysis with FFT. Based on the data collected from experiments, we devised a statistic model for classification.