Subpixel edge extraction in noisy image using a GNAR model

When measuring a part dimension in the condition of machine vision, an accurate and robust boundary position of the part is required. So edge extraction is a fundamental technique in image processing that requires subpixel accuracy. However, traditional subpixel extraction methods are computationally inefficient especially when images are affected by noise. The image data are output of the image collection system; therefore, it can be comprehended according to system analysis that the edge subpixel series is a stochastic sequence which presents the location of the discrete edge spots in image coordinate. Hence, it was proposed to determine the image boundary position by a kind of time series model, a general expression for nonlinear autoregressive model (GNAR model). The models were used to fit these series and the mean term outputs of the models were positions of the effective edges. The experiment results show that describing the edge with stochastic model is in line with the actual law of imaging and the subpixel edge extraction method based on GNAR model has good antinoise capability and high detection accuracy.