Augmented Data and improved noise residual-based CNN for printer source identification

With the classification revolution driven by convolutional neural networks (CNN), plenty of suitable CNN architectures are readily available. However, a limiting aspect of CNN is their hunger for data. Training data can be scarce in many forensics applications, particularly for printer forensics where hard copy pages need to be printed and then digitized using scanners. This paper aims to tackle this problem by using: 1) complementary representations of data and 2) augmented data variations. We derive a simple yet effective noise residual representation of a character image that complements the information contained in the original image. Further, rotated character variations and their monotonic grayscale transformations are used as augmented data. Since such variations are directly linked to printer mechanism, they help improve the overall classification accuracy. Finally, spatial pyramid pooling is used to accommodate characters of all sizes without compromising on a character's spatial information. The proposed method outperforms state-of-the-art CNN based method on a publicly available dataset, and its generic nature allows it to be combined with any other CNN architecture as well. Experiments indicate that for the case of very limited training data availability, the proposed method can achieve 2.5% increase in printer classification accuracy.