On noise reduction for handwritten writer identification

Academic work in identifying writers of handwritten documents has previously focused on clean benchmark datasets: plain white documents with uniform writing instruments. Solutions on this type of data have achieved hit-in-top-10 accuracy rates reaching upwards of 98%. Unfortunately, transferring competitive techniques to handwritten documents with noise is nontrivial, where performance drops by two-thirds. Noise in the context of handwritten documents can manifest itself in many ways, from irrelevant structured additions, e.g., graph paper, to unstructured partial occlusion, e.g. coffee stains and stamps. Additional issues that confound algorithmic writer identification solutions include the use of different writing implement, age, and writing state of mind. The proposed work explores training denoising neural networks to aid in identifying authors of handwritten documents. Our algorithms are trained on existing clean datasets artificially augmented with noise, and we evaluate them on a commissioned dataset, which features a diverse but balanced set of writers, writing implements, and writing substrates (incorporating various types of noise). Using the proposed denoising algorithm, we exceed the state of the art in writer identification of noisy handwritten documents by a significant margin.