Movie Genre in Multi-label Classification Using Semantic Extraction from Only Movie Poster

Abstract

In this paper, we present the framework, SEMPD (the Semantic Extraction of Movie poster based on fundamental of poster Design) for multi-label genre classification in the state of insufficient data, included only movie poster. In order to get manageable semantic tags, we use a combination of fundamental movie poster design in order to understand the result of movie genre. With the collection of a movie poster, we decompose into twelve meaningful features. Then, eight-teen genres are classified by using multi-label prediction algorithms. The results proved that our semantic features perform better than without any information. The experiment has shown that applying Multinomial Naive Bayes with Label Power Set can perform Jaccard score 41.78% compared with baseline 1.11% without any meaningful information. On the other hand, a movie poster can be collected with more manageable features in order to enhance multi-label genre classification.