Concept Analysis in Movie Posters via Convolutional Neural Networks

• Main Authors: Lin, Chun-Ju, 林君儒
• Other Authors: Yeh, Mei-Chen
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/7z66cz

碩士 === 國立臺灣師範大學 === 資訊工程學系 === 105 === In recent years, people have a variety of entertainments; however, watching movies is still the primary choice of many people. Movie posters are playing an important role in advertising a film. People can easily capture the concepts of a poster based on the visual cues it reveals. But, what exactly are the concepts? In this paper, we assume that the design of a movie poster is related to the movie genre; in other words, movies of the same genre may use a similar style in designing the movie posters. We collect movie posters from the IMP Awards website released during 2006 to 2015 as a study case and obtain the genres and keywords of each movie from the IMDb website. We use the Convolutional Neural Network as the main analysis technique, which has shown excellent performances on image recognition, to extract the features (neuron values) of a movie poster. Finally, we analyze the correlation between neuron values and keywords (and emotions), which are considered concepts a movie poster may have. Our study shows that using Convolutional Neural Network for classifying movie posters has a great performance, and the dimension of the Fc7 layer doesn’t affect the classification effectiveness. However, the correlation between neuron values and keywords (and emotions) is not obvious using the analysis approaches proposed in this thesis.