Research Highlights

POSTECH and APGC Startup Company Develop Foreign Language Translation Smart Glasses

2015-10-08 713
 A team comprised of Prof. Daijin Kim (Intelligent Media Lab) and StradVision Inc.,

 

A team comprised of Prof. Daijin Kim (Intelligent Media Lab) and StradVision Inc., a startup company of the Association for POSTECH Grown Companies (APGC) under the leadership of CEO Dr. Bongjin Jeon, developed an automatic text region detection and recognition technique that received an award in the “Robust Reading of Text in Challenging Contexts” category of the International Conference on Document Analysis and Recognition (ICDAR 2015) held in France. These smart glasses automatically detect and translate text into the user’s native language. They can also provide related information on the extracted text.
 
More than a thousand teams from fifty countries participated in the conference including global IT companies and universities. As the most prestigious competition in text detection and recognition, teams compete for the most accurate detection and recognition of text using Google Glass. POSTECH’s team ranked first place in five categories and tied second (without first rank) among seven fields in the competition, implying that the team’s technological power was proven to be the most competitive.
 
Prof. Daijin Kim, Dr. Bongjin Jun, Myung-Chul Sung, and Dr. Hojin Cho also submitted an oral presentation paper titled Scene Text Detection with Robust Character Candidate Extraction Method. In this research, the team uses a modified variant of maximally stable extremal regions (MSER) to accurately extract character candidates in natural scene images. Instead of selecting the most stable region from each local path in the extremal region tree, regions are first grouped according to their similarities and stabilities and the most stable region from each sub-path is selected.
 
After extracting character candidates from the image, the candidate regions are verified using an AdaBoost trained classifier. The remaining regions are further refined using simple heuristics and then grouped for evaluation in public benchmark datasets.
 
The team has worked for more than a decade on developing face detection and recognition. They plan to extend the developed technology to core technologies for smart cars which include the detection and recognition of lanes, cars, pedestrians, traffic signs, and road signs.