A new edge detector based on the statistical vector and neural network is presented in the paper. Firstly, based on the distribution characters of intensities in the neighborhood of an edge pixel, a statistical vector composed of 4 components is proposed. Then through the training with the statistical vector samples calculated from training images, the BP neural network acquires the function of a desired edge detector. Finally, the trained BP neural network is used for edge detection directly. Experiments are carried out with both noisy artificial and natural images. The proposed edge detector proves robust against noise. Besides, both the architecture and training of the BP neural network are simple. Moreover, the proposed edge detector needs no thresholds for conventional edge detection methods.