Abstract: Aiming at the problem that tradition handwritten numeral recognition method can not solve the interference from writing arbitrary,a new handwritten numeral recognition method was proposed based on nmulti-scale features and neural network．Firstly,two structural features of outline and strokes were extracted,and multi-angle structural features were extracted by rotating the datum line．Second,Multi-level grayscale pixel features were extracted by dividing the ima-ge to K sub-layer from the inside out．Thirdly,BP neural network model was build based on the two features．Lastly,new method was used for The MNIST font library,and the prediction precision reached 99.8%．The result shows that new algorithm can effectively reduce the impact of tilt.

Key words: Multi-scale,Handwritten numeral recognition,Multi-angle structural features,Multi-level grayscale pixel features