Latent fingerprints are important evidences used by law enforcement agencies to identify suspects for centuries. However, due to the poor image quality and complex background noise, separating the fingerprint region-of-interest from complex background is a very challenging problem. This paper proposes a new latent fingerprint segmentation method based on Convolutional Neural Networks (ConvNets). The latent fingerprint segmentation problem is formulated as a classification system, in which a set of elaborately designed ConvNets is learned to classify each patch as either fingerprint or background. Considering the spatial correlation between fingerprint patches, we proposed to train the set of ConveNets using multi-sized overlapping patches to utilize complementary information. Then, a score map is calculated based on the classification results to evaluate the possibility of a pixel belonging to the fingerprint foreground. Finally, a segmentation mask is generated by thresholding the score map and used to delineate the latent fingerprint boundary. Experimental results on NIST SD27 latent database demonstrate that the proposed method outperforms the existing benchmarks in terms of both false detection rate (FDR) and overall segmentation accuracy. Thanks to the off-line training and short segmentation running time, the proposed method is applicable to applications such as latent fingerprint matching.