메타정보기반 이미지얼굴표정인식개선

Abstract

Improved Recognition of Facial Expressions in Images based on Meta Information Moon, HyeongJu Advisor : Prof. Shin, JuHyun Ph.D. Department of Software Convergence Engineering Graduate Schoo lof Industrial Technology and Entrepreneurship, Chosun University Due to the impact of the pandemic and the development of ICT technology in the era of the 4th industrial revolution, the use of non-face-to-face /unmanned systems is expanding, human facial expression recognition plays an important role in interpersonal relationships and non-verbal communication, and research on emotion recognition through human facial expressions has been ongoing for a long time. Recently, various artificial intelligence technologies are achieving higher performance in various applications than conventional methods that do not utilize artificial intelligence [1] In this regard, many researchers are proposing machine learning and deep learning models to improve image facial expression emotion recognition, and most emotion recognition studies are solving problems in various fields by using deep learning models after extracting various facial expression features of facial images [2]. Therefore, this study aims to improve image facial expression recognition by utilizing meta-information of Korean complex image data of AI Hub (AI-Hub), an integrated platform operated by the Korea Intelligence Information Society Agency. Existing facial expression emotion recognition studies utilize a large amount of data to improve accuracy. However, learning a large amount of data requires high computing power and requires the experimental environment and the collection of a large amount of data. To improve this, we propose an emotion recognition method using age and gender, which are meta information of image data, to improve emotion recognition performance even with a small amount of data. The face was detected through the Yolo v2 Face model in the original image data, and a series of emotion classification processes were proposed to classify emotions after age and gender classification with the VGG16 model using the meta information of the image data, and the effect was evaluated. Based on the meta-information included in the image data, data classification was conducted by age (10-50s) and gender (male and female), and then 10 emotion recognition models were created using the EfficientNet model with the best emotion recognition performance among VGG, ResNet, and EfficientNet models. Next, it was confirmed that the accuracy performance of the learned model improved by classifying meta-information-based data and comparing and evaluating the performance of the learned model with the whole data that was not classified. The total data that is not classified is a total of 70,000 sheets, and the data of the proposed model is 7,000 sheets, which is 1/10 of the total data. For this reason, the significance of this study can be found in that it proposes a method to improve performance with low computing power when recognizing the proposed meta-information-based image facial expression.