The fusion recognition of multi-band images can extend the application range of recognition systems. A fusion method based on convolutional neural networks (CNN) was explored and designed in this paper. Based on the AlexNet network model， it was extracted that the ship target features of three wave band images concurrently in visible light， Middle Wave Infrared (MWIR) and Long Wave Infrared (LWIR) bands. Then， it performed the feature selection for concatenated three-band eigenvectors by using the mutual information method and determines the dimensions of fusion eigenvectors according to sorting the importance of concatenated feature eigenvectors. Finally， three fusion methods named as Early fusion， Middle fusion and Late fusion were used to verify respectively the effectiveness of the proposed algorithm according to the features extracted from different levels. An available ship target dataset in three bands containing 6 categories of targets and more than 5 000 images was established for our experimental verification. The results show that the recognition rate from Middle fusion reaches 84.5%. Compared with Early Fusion and Late Fusion， it increases by 8% and 12%. Moreover， the recognition rates of all three fusion methods have been improve significantly as compared to that of the single band recognitions at the same application scene.