Neural Networks Applied to Recognition of CCD Camera Images where a 3-Digit Number is Stamped on the Surface of Chip Resistors

We describe Back-Propagation neural networks (BP model) ['I implemented into a visual verification system of numbers stamped on the surface of chip resistors. A 3-digit number is stamped whitely on the rectangular black area, that is on the su@ace of each chip resistor, bordered by a white frame. The number and a fragment of the frame are imaged from a CCD camera. Firstly, the number is circumscribed by a rectangular frame for segmentation in each CCD camera image (this is a segmentation module). Secondly, while a narrow slit window is scanning across the partial image inside the segmented rectangular area, each digit is classified one after another concurrently with separation from its neighbors (this is a recognition module). Both modules are developed by training BP models, Position Network and Recognition Network. Experimental results show that the proposed method could correctly recognize the number even if digits were stamped poorly and imaged with extraneous substances.