효과적인 감정 분석을 위한 저 레벨 영상 특징(low-level image feature)추출

초록

Sentiment analysis has received considerable critical attention in machine learning, artificial intelligence, human–computer interface, etc. In these fields, many studies have analyzed emotions using images, audio, and bio-signals as features. Among them, those that utilize image features are the most typical for emotion recognition. There are two types of image features: high-level and low-level. Low-level features are more robust than high-level in sentiment analysis. Therefore, in this paper, we investigated the critical features for effective sentiment analysis in low-level image features. For an objective performance evaluation, we utilized the International Affective Picture System dataset for training and testing. We applied the iterative Han and Cha’s feature selection/extraction algorithms and used a multilayer perceptron classification. We also carried out cross-validation by replacing features. Our evaluation items consisted of two elements: accuracy and computational time. According to our results, we were able to find the critical features in sentiment analysis and our method proved more competitive compared with existing algorithms in terms of accuracy and operation time.

키워드