Fingerprints have been widely used in forensics and law enforcement application for more than a century. The matching performance of rolled and plain fingerprints, which are captured in an attended mode, is extremely high. However, the matching of latent fingerprints, which are lifted from the surfaces of objects inadvertently touched or handled by a person typically at crime scenes, is still a challenging problem due to complex background noise, poor ridge quality and overlapping structured noise in latent images. Inspired by the successful application of dictionaries in number of signal processing tasks, such as image denoising, classification and face recognition, we investigate dictionaries for some challenging problems in fingerprint image analysis. Two kinds of dictionaries, orientation patch dictionary and ridge structure dictionary, are investigated. Orientation patch dictionary, which contains only orientation information in patches, is proposed to estimate orientation field for latent fingerprint enhancement on manually markup region of interest (ROI). Ridge structure dictionary, which contains ridge and valley patterns, is proposed for segmentation (ROI detection) and orientation and frequency fields estimation for latent fingerprints. Experimental results on two latent databases, NIST SD27 and WVU DB, demonstrate dictionary‐based methods are able to improve latent fingerprint matching performance. A specific ridge structure dictionary around minutiae is proposed to estimate minutiae quality, which is an important factor for latent fingerprint quality estimation. These two kinds of dictionaries are also able to improve fingerprint reconstruction performance from minutiae set, in which orientation patch dictionary is used to reconstruct orientation field and ridge structure dictionary is used to reconstruct ridge and valley pattern.