This papers aims to improve the ship detection precision in optical remote sensing images and realize sophisticated classification of ships without need to conduct complicated pre-processing step and manual feature extraction. For input detection images, selective search method was used to generate candidate ship regions. The parameters of convolutional neural network were obtained by supervised training with labeled training data. Then the trained convolutional network was utilized to classify the candidate regions. The classification results of ship candidate regions were used to locate the ship and identify the category of ship. Compared with two existing detection methods, the experimental results indicated that the proposed convolutional neural network based method could efficiently improve the detection precision and achieve an average detection precision of 93.3%. In conclusion, the method can detect and classify ships simultaneously without complicated pre-processing step. And it also can improve the efficiency of ship detection precision as well.