摘要 |
PURPOSE: To accurately inspect the various kinds of image patterns with no alignment of each pattern in position, and without storing any reference image pattern. CONSTITUTION: An inspection object signal OD is developed into the image elements of 7×7 in a two-dimensional manner by the use of line memories LM1 through LM6 and shift register groups MR, and the values of peripheral image elements 01 through 040 excluding the image elements of 3×3 in the vicinity of the center out of the image elements of 7×7 are inputted to an estimated net-work 100. The estimated network 100 is formed out of neural networks, and has made a study in advance using a reference image data. Namely, parameters are so set by a learning operation circuit 102 that the difference between the value of a central image element 00 and the value of an output Dout corresponding to the aforesaid central image element becomes smaller. The output value Dout obtained with the inspection object signal OD inputted after a study has been made, is binarized by a judgement circuit 104, and then outputted as a non-defective estimated data PD. The defect of a tested article is detected based on the comparison of the non-defective estimated data with the inspection image data OD. |