Abstract: It is the key technology for fast and accurate fingerprint classification to accelerate target fingerprint search in large fingerprint identification systems. There still exist some problems such as complex operation process, numerous parameters, large scale data and inadequate use of the fingerprint information. It’s more critical and representative for the features to be extracted by the deep layer of neural networks, namely, some information in the shallow layer may be ignored. Therefore, in this paper, the lightweight multi-feature fusion fingerprint classification algorithm is presented. Firstly, the fingerprint images are trained by the lightweight Finger-SqueezeNet. Simultaneously, after the look-up-table method is used to obtain the refined map of the fingerprint, the improved distribution summation gradient method is used to get the Region Of Interest(ROI) image of the corresponding map, which is merged with the extracted feature map in the deep layer of the network. Since the deep network receives the accurate trend information of the fingerprint lines in the shallow layer, the sensitivity of the network to the pattern is enhanced. Finally, the Maxout activation function is used to activate the features extracted by the network. Experimental results show that the network structure of the five Fingerfire modules is optimal for fingerprint classification, and the network model after feature fusion can reach 96.81%. Moreover, the test result obtained by a single fingerprint test method reaches 94.57%. Compared with the same type and verification method, the fingerprint classification model in this paper can still obtain higher accuracy while greatly reducing network parameters. In addition, the classification of the whorl fingerprints has a good performance among the five categories, but the arch and the tented arch are relatively poor, which is because of the 17.5% fuzzy labels of these two classes fingerprint. It is worth mentioning that two kinds of fuzzy fingerprints haven’t been mixed as one class to improve the result, nor any of them has been refused to be recognized, namely, it’s 0 rejection rate. It can be seen from the final results that the model has strong generalization ability and high stability for the classification of fingerprint with different quality. The method can make full use of fingerprint information by compensation of the ROI image, and the lightweight algorithm provides theoretical support for the extension of the fingerprint classification model to the mobile end.