To cope with the problem that the traditional fine feature extraction methods for identifying communication transmitters suffer from the lack of the labeled samples in real complex electromagnetic environment, an efficient fine feature extraction method, called locally neighborhood preserving regularized semi-supervised discriminant analysis, is proposed for communication transmitter recognition. Based on the bispectrum estimation, manifold structure information is incorporated into the linear discriminant model by unlabeled samples, which extends the linear discriminant analysis to the semi-supervised learning. Extensive experiments on the real-world database sampled from different FM communication radios with the same model, manufacturer, manufacturing lot, and work pattern demonstrate that the proposed method can obtain better recognition performance.