Latent fingerprints are (un)intentional finger skin impressions left as ridge patterns at crime scenes. The significant challenge in latent fingerprint segmentation is extracting complex, multiple, noisy foreground fingermarks while maintaining the performance of the system. The work presented in this paper provides a method to extract fingerprints from the latent fingerprint images dataset (IIIT-D) using a stack of convolutional auto-encoders. The idea is to early detect the structure of interest from the image using a color-based mask. These structures are divided into equal-sized patches and classification of these patches into fingermark or background class-labeling is achieved using staked convolutional autoencoders. To establish stable layered architecture and an optimal amount of information in patches as input to these layers, the impact of different patch-size is analyzed on various stacks of the layered architecture of the underlying deep neural network. Reduced feature learning of an autoencoder and pre- trained convolutional neural network improves the patch classification accuracy thereby increasing segmentation accuracy.