Biometric systems are widely deployed in governmental, military and civilian applications. There are a multitude of sensors and matching algorithms available from different vendors. This creates a competitive market for these products, which is good for the consumers but emphasizes the importance of interoperability. Interoperability in fingerprint recognition is the ability of a system to work with multiple fingerprint scanning devices. Assuming that the same sensor or vendor will always be available during the lifetime of an automatic fingerprint recognition system is unrealistic. Typical variations induced by fingerprint sensor diversity include image resolution, scanning area, gray levels, etc. Such variations can impact (i) the quality of the extracted features and (ii) cross-device matching performance. In order to enhance interoperability, previous research has proposed a variety of fingerprint feature representations as well as a classification scheme to improve match rates across devices; however, implemented systems are not good enough to be operative. In this work, we propose a learning-based compensation scheme based on features derived from the discrete wavelet transform (DWT) and binarized statistical image features (BSIF) of captured fingerprint images. In particular we are interested in DWT for its capability to preserve spatial information of an image when performing frequency analysis while BSIF has shown to be effective in texture recognition tasks as a local descriptor. Experiments are carried out on a data set consisting of fingerprints obtained from 494 users acquired using four different optical devices. Results show reduced error rates compared to the baseline as well as improved performance compared to previous research.