This paper proposes a model-based approach to extract natural strokes in handwritten Chinese character images. We model the distortion between an input stroke and its counterpart in the character model as the result of an affine transformation followed an uncertainty distribution. Based on this modeling, a statistical method is presented to measure similarities between strokes or characters, and a probabilistic relaxation process is put forward to assign line segments in the skeleton of the input character to model strokes transformed by detected affine transformation. All line segments assigned to a model stroke constitute the corresponding input stroke. For achieving high robustness, affine transformation estimation and stroke extraction are performed alternately until the similarity between the input character and the character model is maximized. The proposed stroke extraction and matching method was applied to off-line handwritten Chinese character recognition, whose effectiveness is confirmed by the experimental results.