Sparsity-based super-resolution for offline handwriting recognition
Shiv Naga Prasad Vitaladevuni, Huaigu Cao, David B. Belanger, Krishna Subramanian, Rohit Prasad, Prem Natarajan · 2011
We present a sparsity-based approach to super-resolution for handwritten document images, and demonstrate that it improves handwriting recognition accuracy. Given high resolution training images, low and high resolution dictionaries are constructed by extracting patches. The low resolution patches are adapted to expected distortions in the out-ofdomain test data using image filters. The intuition is that the low-resolution patches would match with artifacts in the test images and the pristine high-resolution patches would be back-projected to get a high-resolution version of test image. Patches from test images are projected onto the lowresolution dictionary under sparsity constraints. The projections coefficients are used to back-project high-resolution dictionary elements for super-resolution. Our experiments indicate that this super-resolution produces substantial improvements in handwriting recognition over bicubic. An important feature is the use of a separate, out-of-domain high resolution dataset for learning the dictionary and adapting it due to the unavailability of high resolution versions of the test data.