In order to improve the accuracy and resolution in edge location, the systematic error in Canny’s sub-pixel edge detector is studied, as well as a method to compensate it is presented in this paper. The digital edge is derived from an ideal step edge through imaging and sampling. Because the high frequency part exceeding the cut off frequency of normal optical systems is not zero, for a given image sensor such as a CCD, there will be some certain aliasing from the insufficient sampling, which results in the systematic error in sub-pixel edge location. This error is periodic in each pixel and is affected by the aberration of optical system and noise. A compensation method is presented based on Multilayer-Perceptron Artificial Neural Network (MLP-ANN). The training data set is constructed by moving and tilting a straight edge randomly, the output of MLP-ANN is the compensation for the systematic error. It is shown that in high signal noise ratio images, systematic errors are main sources of edge location error. The standard deviation of location error for a straight edge was 0.019 pixels in our setup, and after compensation, the standard deviation decreased to 0.010 pixels, this can meet the requirement for most practical engineering measurements.