(A) hierarchical probabilistic framework for scene text separation

Abstract

Automated image understanding could be useful so that extracted information could be used in further ways. Since text has been the most significant information medium for the mankind, an automatic system that reads text has been one of the most popular research targets. However, when text reading is applied to $\textit{scene images}$ taken by digital cameras, we meet a different class of text from traditional ones, aka $\textit{scene text}$, which shows highly various characteristics. A common process to recognize text from an image is first to separate text pixels from background, and then recognize the text-only image. Separating text pixels from scene images remains an open problem, since scene images have much variation and properties of text such as position, size, and color are not constrained. In this research, we cast the scene text separation problem into probabilistic labeling in which we yield the labels which maximize its conditional probability given the image. Based on the hierarchical model of scene text with four layers as image - text line - stroke - label, we provide pixel-oriented representation of objects in each layer to build a random field for describing possibilities of objects in the layer. According to the hierarchy, the proposed hierarchical framework decomposes the labeling problem into three sub-problems as one about text lines, one about strokes, and one about labels, where two formers are described using the Kernel Ridge Regression (KRR), and the latter using the Conditional Random Field (CRF). The optimal labels are identified through a stochastic gradient method. Our framework is built learnable so that the parameters in it can be trained to improve the performance. The experimental results showed the promising performance of our framework.