Rotation scale and translation invariant (RST-invariant) feature extraction methods play a critical role in optical Chinese character recognition. However, vast majority of them either has a strict constraint of input image, or has a poor performance on discriminating similar Chinese characters. In this paper, a novel method called Sector Projection Fourier Descriptor (SP-FD) is proposed. SP-FD is a region-based Fourier descriptor which contains two stages. Firstly, the input character image is transformed into polar space through sector-projection, which generates a periodic function named sector-projection shape signature. Secondly, feature vector is obtained through 1-D Fourier transformation on the shape signature. In our method, the input image does not require to be normalized as a precondition, and the shape information of original character is not assumed to be available neither. Moreover, the internal structure of character in the circular direction is exploited. The experimental results show the proposed method can extract the RST-invariant feature effectively, and outperforms the typical algorithms on discriminating similar Chinese characters.