Studying On Emotional Modeling For Visual Media

Visual media contains a wealth of information, and is widely infiltrate people's work and life now, which produce a subtle influence for people on Physiological and psychological. Therefore, Analysis and model building on relationship between visual media and people's emotional reactions, have become an important interdisciplinary subject among psychology, physiology and computer science at present.This thesis aims to study the establishment of an integrated model across the video space, emotional space and EEG space. The main works include the following three aspects.First of all, we conduct affective analysis on international standards affective audio library IADS and international standards affective image library IAPS. For IADS, we carry out analysis with various classification methods and make comparisons with the results in feasibility and effectiveness of these methods. For IAPS, in addition to the analysis of classification and comparison just like for the IADS, we also conduct emotional dimension rules mining by using images' low-level features, and acquired some useful rules for reference.Secondly, we have established a video clip library for emotion stimulating, and select part of them to elicit subjects' emotion changes and physiology reactions. Elicited EEG signals and subjects' self emotional evaluations are both recorded, and then a database, which contents across video, EEG and emotional dimension 3 kind of spaces, is built. Based on this self-built database, we first establish two one-way mapping model for video-emotion and EEG-emotion respectively, and then make research on the two-way mapping model between video and elicited EEG. At last, to apply our emotion research results into usual life, fuzzy clustering is carried out for emotional dimension, and relations between emotional dimension and emotional categories are acquired. As a result, either exploiting stimulate video clips or EEG signals can get a mapping emotion categories results finally. In addition, the direct relations between video and the induced EEG signals are established partly, which is rarely studied in this field at present.In the end, due to the time length of video clips, It will take too much time for subjects to evaluate the emotion induced by video clips, and it is easy to make subjects feel tired. This kind of user fatigue become a great obstacle for more reliable affective video analysis with the expansion of samples. To solve the problem of lager number size labeling, we try to complete the automatic labeling work for unlabeled video by utilizing semi-supervised learning methods, and list the analysis results.Through the contents of the above three aspects, a relatively comprehensive affective analysis for visual media is completed. Not only confirm the exist of relations between stimulative video clips and elicited EEG signals, but also establish several specific mapping models, and accomplish the initial model mining works involving video space, psychology space and physiology space.