This paper proposes an incremental scheme to identify the neurons in electronic microscope image. It first computes the probability of each pixel being cell membrane and then determines the image areas of the neurons. Our contributions also include novel combined multi-scale image features that are computationally efficient and provide strong discrimination ability for differentiating membrane pixels from other pixels. Experiment results show that the proposed system offers much higher speed than the compared state-of-the-art methods while produces comparably high accuracy.