Decoding Emotion From FMRI Based On Machine Learning

Images contain very rich emotion meanings. Image emotion meaning recognition is a hot spot in many research fields which include digital image processing, artificial emotion, machine learning and cognitive psychology. Image emotion is an active, internal and implicit emotional reflection of the man when he or she watches the image. Most of the existing literature are selected based on low-level features of image itself, through machine learning methods, and establish the relationship between low-level features and semantics of image emotion. The emotional meaning annotated image library is the key point of this kind of researches. However, in the most of existing image libraries, the emotional meanings are manually annotated based on the statistics. The manually annotated image library is very expensive and hardly reflects the internal and implicit emotion of human being.Functional magnetic resonance imaging is an imaging technology, mainly by measuring the blood oxygen level dependent signal activation of the brain. With fMRI study of the emotional process of the application, the researchers found that different emotions will have different activation patterns. In this paper, the fMRI is used to observe the brain reactions to the images and the Machine Learning method is introduced to establish the map between brain activation patterns and the human emotions. After practice, a classifier can be gotten to automatically annotate the emotional meanings on the new images based on the map.This research is a part of the project Images Annotation Based On fMRI for Personalized Emotion and Ontology Library research that is supported by the national natural science foundation of china under grant No.60970059 and is the exploratory research. This project majorly researches on the different brain activation patterns reflected by different human emotions, select the appropriate feature selection method to improve the accuracy of classification, to achieve the type of emotional interpretation of fMRI data. Main researches are as follows:1. The different brain active areas reflected by different human emotions are positioned. Image emotional resonance design experiments using event-related design. An image emotion fMRI experiment is designed and the emotional meaning clear (non emotional meaning ambiguous) pictures that include scene, animals and people are used in the experiment. Using Statistical Parametric Mapping (SPM) software to obtain fMRI data were analyzed by the f-test comparing means, access to positive, neutral, negative contrast the three conditions activated brain areas. Experimental results show that the Extra-Nuclear, Superior Frontal Gyrus, Precentral Gyrus, Middle Frontal Gyrus, Superior Temporal Gyrus, Parietal Lobe, Occipital Lobe and other brain regions were significantly activated, the activation of brain areas have been reported in the literature and consistent.2. FMRI data with small sample size, high dimensional features. Comparison of classification performance, the study found, use relatively simple linear classifier for classification of fMRI data, the level of classification accuracy rate of more than random, with an average accuracy close to the subjects involved in the subjective forecast, and has faster processing speed. The results also show that the linear classification of fMRI data can interpret its underlying emotional state.3. Classifier algorithm is the key to choose the right features. In this study, voxels as features, feature reduction through feature extraction and feature selection in two ways:First, feature extraction, select the active zone voxels as features; Then, feature selection, based on single-voxel accuracy or activation ranking voxels, and voxels from which to choose the best. This method effectively improve the classification accuracy of up to 90%.