著者
西銘 大喜 遠藤 聡志 當間 愛晃 山田 孝治 赤嶺 有平
出版者
一般社団法人 人工知能学会
雑誌
人工知能学会論文誌 (ISSN:13460714)
巻号頁・発行日
vol.32, no.5, pp.F-H34_1-8, 2017-09-01 (Released:2017-09-01)
参考文献数
20
被引用文献数
6

Facial expressions play an important role in communication as much as words. In facial expression recognition by human, it is difficult to uniquely judge, because facial expression has the sway of recognition by individual difference and subjective recognition. Therefore, it is difficult to evaluate the reliability of the result from recognition accuracy alone, and the analysis for explaining the result and feature learned by Convolutional Neural Networks (CNN) will be considered important. In this study, we carried out the facial expression recognition from facial expression images using CNN. In addition, we analysed CNN for understanding learned features and prediction results. Emotions we focused on are “happiness”, “sadness”, “surprise”, “anger”, “disgust”, “fear” and “neutral”. As a result, using 32286 facial expression images, have obtained an emotion recognition score of about 57%; for two emotions (Happiness, Surprise) the recognition score exceeded 70%, but Anger and Fear was less than 50%. In the analysis of CNN, we focused on the learning process, input and intermediate layer. Analysis of the learning progress confirmed that increased data can be recognised in the following order “happiness”, “surprise”, “neutral”, “anger”, “disgust”, “sadness” and “fear”. From the analysis result of the input and intermediate layer, we confirmed that the feature of the eyes and mouth strongly influence the facial expression recognition, and intermediate layer neurons had active patterns corresponding to facial expressions, and also these activate patterns do not respond to partial features of facial expressions. From these results, we concluded that CNN has learned the partial features of eyes and mouth from input, and recognise the facial expression using hidden layer units having the area corresponding to each facial expression.
著者
髙嶺 潮 遠藤 聡志 Kolodziejczyk Jakub 西銘 大喜
出版者
一般社団法人 人工知能学会
雑誌
人工知能学会全国大会論文集
巻号頁・発行日
vol.2019, pp.4I2J204, 2019

<p>機械が現実世界の空間情報を獲得するための重要な手段の一つが単眼深度推定である。人間は深度推定に使用できる情報の種類を増やし、問題領域を分割することで精度の高い深度推定を実現している。これを受け、深度以外の情報をRGB画像から獲得することによって単眼深度推定を改善しようとする試みが幾つか存在する。Semanticラベルを用いた実験では、解釈可能な意味の種類が多いラベルが入力画像の幅を制限することがわかり、人間の主観によって定義された情報の欠点を浮き彫りにした。対して、深度勾配を扱った実験は、推定結果の外れ値の削減に大きく貢献している。これらの結果は、数値的に定義可能なオブジェクト情報が、人間が深度推定を行う際に獲得する冗長性の再現に繋がることを示唆している。本研究は、物体の前後関係情報の推定を行うことで深度推定を分類問題の分野に落とし込み、単眼深度推定の精度向上を狙うものである。Multi-Scale Modelを用いた対照実験により、重なり情報の有効性が証明された。</p>
著者
西銘 大喜 遠藤 聡志 當間 愛晃 山田 孝治 赤嶺 有平
出版者
一般社団法人 人工知能学会
雑誌
人工知能学会全国大会論文集 第29回全国大会(2015)
巻号頁・発行日
pp.3L43, 2015 (Released:2018-07-30)

本研究では、ディープニューラルネットワーク(DNN)と表情画像を用いて喜び、悲しみ、驚き、怒り、嫌悪、恐怖、無表情の7感情の推定を行う。6層のDNNでは、65%の認識精度が確認され精度が良い表情は喜び、驚き、無表情の3種類であった。人間との比較実験でも同様の結果が確認され、画像だけの情報では悲しみ、怒り、嫌悪、恐怖の4感情は認識が難しいと考えられる。より多層のDNNを用いた実験と共に報告する。
著者
西銘 大喜 遠藤 聡志 當間 愛晃 山田 孝治 赤嶺 有平
出版者
一般社団法人 人工知能学会
雑誌
人工知能学会論文誌
巻号頁・発行日
vol.32, no.5, pp.F-H34_1-8, 2017
被引用文献数
6

<p>Facial expressions play an important role in communication as much as words. In facial expression recognition by human, it is difficult to uniquely judge, because facial expression has the sway of recognition by individual difference and subjective recognition. Therefore, it is difficult to evaluate the reliability of the result from recognition accuracy alone, and the analysis for explaining the result and feature learned by Convolutional Neural Networks (CNN) will be considered important. In this study, we carried out the facial expression recognition from facial expression images using CNN. In addition, we analysed CNN for understanding learned features and prediction results. Emotions we focused on are "happiness", "sadness", "surprise", "anger", "disgust", "fear" and "neutral". As a result, using 32286 facial expression images, have obtained an emotion recognition score of about 57%; for two emotions (Happiness, Surprise) the recognition score exceeded 70%, but Anger and Fear was less than 50%. In the analysis of CNN, we focused on the learning process, input and intermediate layer. Analysis of the learning progress confirmed that increased data can be recognised in the following order "happiness", "surprise", "neutral", "anger", "disgust", "sadness" and "fear". From the analysis result of the input and intermediate layer, we confirmed that the feature of the eyes and mouth strongly influence the facial expression recognition, and intermediate layer neurons had active patterns corresponding to facial expressions, and also these activate patterns do not respond to partial features of facial expressions. From these results, we concluded that CNN has learned the partial features of eyes and mouth from input, and recognise the facial expression using hidden layer units having the area corresponding to each facial expression.</p>