Securing core source technologies in the fields of visual intelligence and content visual search
In the future, it will be used in public services and intelligent content fields such as CCTV, black box in-depth analysis, and building search.

A domestic research team has been recognized for its image injection detection and average detection accuracy, which are essential for the development of artificial intelligence, at an international image recognition competition.

ETRI (Electronics and Telecommunications Research Institute) announced on the 27th that it achieved 2nd place in detection performance by object type and 3rd place in average detection accuracy by competing with joint teams of companies and universities from around the world in the object detection field of the International League of Science and Technology Recognition Competition (ILSVRC, ImageNet) held at the Hawaii Convention Center.

The object detection field finds the type and location of objects among 200 objects (65,500 photos), and evaluates performance by detection performance and average detection accuracy (mAP) by object type. ETRI, with the participation of the 'Deep View' team and the 'Content Visual Browsing' team, focused on designing and training a network that searches for the type and location of objects based on deep learning technology to improve detection performance and accuracy.

As a result, ETRI's 'Deep View' team achieved 2nd place in terms of detection performance by object type. The 'Deep View' team showed the best performance for 10 objects. A total of 14 organizations participated, and only 2 teams recorded the best performance for more than 10 object types.

The 'Content Visual Browsing' team, together with Professor Shin Jin-woo's team at KAIST, achieved 3rd place with a mean detection accuracy (mAP) of 0.61. mAP is an indicator of the detection accuracy of 200 objects and represents the comprehensive level in the field of object detection.

The research team participated in the object classification and image localization fields of this competition last year, and achieved 5th place in the image object localization field with an error rate of 9.92%, and in the object classification field with an error rate of 3.25%.

Meanwhile, on the 21st, the research team also won second place in the International Low Power Image Recognition Competition (LPIRC) held separately on-site. This competition, held in the field of 'Reboot Computing' by the Institute of Electrical and Electronics Engineers (IEEE), is a competition that measures detection accuracy (mAP) and power consumption (Power) during detection.