Generative Learning for Character Recognition of Uneven Lighting

Recently, development of a text input system based on a text recognizer with a digital camera or a camera phone is highly demanded. However, character images captured by a camera are often degraded by low resolution, blurring, uneven lighting, camera shake and so on. A promising method to prevent decrease in recognition performance is generative learning. It enhances the recognition performance by creating various artificially degraded images that can be obtained in real scenes from degradation-free images. The created images are used for learning together with the degradation-free images, and compensate for lack of actual degraded images. The degradation processes should be well analyzed to emulate them as real as possible. Ishida et al. proposed degradation models of low resolution and blurring[1]. In this paper, we propose a degradation model of uneven lighting and experimentally confirm its effectiveness.