Classifier design of on-line surface defect inspection system for plates

Classifiers based on BP (back propagation) networks, LVQ1 (learning vector quantization) networks and LVQ2 networks were studied and applied to classification of surface defects of plates. Surface defect images of plates were online captured from production line, and each image was divided into several sub-images of size 64 × 64. Fast Fourier Transform (FFT) was used to get amplitude spectrums of each sub-image. Gray levels of pixels in central region of each amplitude spectrum were input to classifiers based on BP networks, LVQ1 networks and LVQ2 networks as features. Experiments show that classification results of LVQ2 networks are satisfactory.