Most pornographic image detecting systems are based on the results of skin detection. They extract the low level features in the image and obtain high positive detecting rate and high false positive detecting rate as well. The paper proposes to extract local features to obtain high level semantic knowledge. The system detects the salient points in the image and describes the local region's form around the points using SIFT descriptor. Look on these descriptors as visual words, then the pornographic information can be detected based on the words found. Only combining the Bayesian formula with the traditional detecting system, the algorithm takes from the false positive detecting rate remarkably when the true positive rate is kept as before.