User-dependent Recognition About Affective Classification Of Images

The present research in field of affective computing mainly devotes to the affection itself, such as affective recognition or affective regulation. But few researches turn to the selecting of affective material, which ware used to induce the affections or to regulate the affections. Now, expert evaluating is the most common way to get the affective material. This way did not fully considered the users' difference and was pretty much a waste of time. This thesis explored the issue of user-dependent recognition about affective classification of images to preliminarily solve the selection problem of affective material. The IAPS images were used in this thesis to induce the subjects' affections.This thesis aims at building a model about the relation of images and the affective classifications(negativity, neutral and positivity) that somebody has assessed it, where the images are expressed by their feature vectors. So that the affective classification of a foreign image can be deduced for the above subject by the built model, which can also be used in the selection of affective images. The properties of this model can be reflected by recognition rates.The thesis launch from four aspects:First, the features of the images and the way to extract them were introduced. According to the relevant studies on visual psychology, the low-level features of images were selected, which were expressed in the way of vectors. At last, a feature vector of 26 dimensions was obtained for an affective image. Second, experiment to evaluate the affective classification of images. Eight sets of effective data were acquired from 10 subjects when they evaluate IAPS images into negative or neutral or positive images. The 8 sets of data correspond to the eight ways to partition the 60 images that divided by 8 subjects. Third, the establishment and training of recognition model about affective classification of images. The BP neural network was used to build the model. For every subject, part of images form his or her segmentation was used to train the already-built network. For the different setup of the training sets(imbalance data and balance data), 2*8 models of 8 subjects were obtained. Fourth, model verification. The recognition rate which can be calculated from the comparison of the result that was evaluated by the corresponding subject and the result of model was used to reflect the properties of a model. And the result of model can be obtained through inputting features vectors of the rest images into the model.The average recognition rate of the 8 models that were obtained under imbalanced training sets achieved 66.83% while the average rate of the other 8 models obtained under balanced training sets achieved 69.87%. Further, a comparative analysis was done about the two kinds of models. The results show that not only the average accuracy is higher when the balanced-training-sets be used to train the BP network, but also the recognition rates are more evenly between different affective classifications. The standard deviations about the average recognition rates of the three affective classifications are sbalance=0.07 and simbalance=0.09. Therefore, the recognizer model is superior which is trained by balanced training sets.The result of this research shows that it is effective to build the BP recognition model. That's to say, the method of user-dependent recognition about affective classifications of images is feasible. So this thesis provides a way for selecting affective images, and a new thought for affective material.