サッカードに駆動された「傷の気付き」処理

抄録

In this paper, we discuss a few issues about “KIZKI” processing which is useful for surface inspection of raw material. The “KIZKI” processing was developed based on knowledge about peripheral vision and involuntary eye movement. This processing suppresses a background pattern of an inspection image and detects minute defects that are not easily found. The “KIZKI” processing has contributed to the automation of various visual inspections. On the other hand, several issues have been identified during this period. In this study, we bring up two problems about “KIZKI” processing. First, the previous “KIZKI” processing attenuates the output for smaller and weaker defect signals in the presence of multiple defects of different size and intensity. Second, the “KIZKI” processing features only local contrast, so there is a concern that it may detect normal pattern that have the same level of signal as defects. To solve these problems, we propose two types of global algorithm based on knowledge about saccade and perceptual organization. Specifically, the “KIZKI” processing is applied to multiple local regions of an inspection image in turn to detect defect candidate images. This processing is like the way a saccade shifts the gazing points. By limiting the processing area, defects of different size and intensity can be detected with similar output signal. In addition, if the candidate images are organized based on placement or shape features, the output signal is suppressed. This processing suppresses the detection of normal patterns. And this processing is like the perceptual organization. The evaluation experiment showed that the proposed method showed could solve the above two problems.