Amazon Researchers Utilize Artificial Intelligence to Enhance Text Recognition Capabilities

2019-12-29

Optical Character Recognition (OCR) is the process of converting handwritten or printed text images into machine-readable text. This scientific field dates back to the early 1970s, but algorithms have long struggled to recognize characters that are not parallel to the horizontal plane. To address this issue, researchers at Amazon have developed a "text box" detector for curved text in natural images. In a paper describing their work, the authors claim that their method achieves state-of-the-art results on a popular OCR benchmark. The paper explains that scene text is typically divided into two consecutive tasks: text detection and text recognition. The first method involves localizing characters, words, and lines using contextual clues, while the second method transcribes their content. However, curved optical text poses challenges due to its inherent deformability, as well as variations in viewing angles and arbitrary fonts. The team's solution is a "tube" representation of a text reference frame, which captures most of the variability and leverages the fact that target texts are usually composed of similarly sized strings. It is a mathematical function that can train a machine learning scene text detector. Unlike traditional methods that use easily overlapping rectangles and quadrilaterals to capture text information, this approach is different. The researchers evaluated the performance of the "text box" on CTW-1500, a dataset consisting of 1,500 images collected from natural scenes and image libraries, with over 10,000 text instances, each image containing at least one curved instance. They report that the "text box" achieved industry-leading results on CTW-1500, with an accuracy of 83.65%, compared to 75.6% accuracy achieved by recent similar methods. The authors of the paper write, "Modeling the central axis and average radius of an instance... allows us to capture information about the entire instance. On datasets composed of single-word instances, such as Total-Text, our model achieves state-of-the-art performance. On datasets with line-level annotations, such as CTW-1500, our model is better able to capture text information." If this research is implemented in the future, it will undoubtedly be a boon for businesses heavily reliant on OCR. Additionally, it has the potential to further drive the advancement of paperless offices.

RECENT AI NEWS

RECENT AI TOOLS