In this paper, we propose a method for personal identification using facial features and context information. The method can overcome a problem of partially occluded images. In the proposal method, facial similarity is calculated by using CLAFIC method from the full image of face and the parts of face image, such as the image of eyes. Moreover, we focus on clothes of each target person as the context information. In this method, the similarity of clothes is calculated by using four features. Then, from the two similarities, the person in the input image can be identified. In the experiments, it can be confirmed that proposal method is better than the method using only the face image similarity.
Lip motion features such as lip width and lip length provide important information to identify individuals or commands. Previous study has proposed recognition system that used L*a*b color space to acquire lip information. However, the interface using lip motion features has not achieved automatic detection of speech activity. Therefore, we propose a method to detect speech frames using color information of lip region and features of lip motion. The proposed method has three steps. First, it extracts the lip motion features on the basis of both position and shapes of lip in each frame of facial images. Second, it judges the existence of oral slit in each frame of lip images by color information of lip vertical lines. Third, it detects differences of lip length between three frames: the object frame which has oral slit and two previous frames. The experimental results for five persons show that the proposed method is able to detect speech frames with an accuracy of 99.2 percent.
This paper presents Facial Expression Spatial Charts (FESCs) as a new framework to describe individual facial expression spaces, particularly addressing the dynamic diversity of facial expressions that appear as an exclamation or emotion. The FESCs are created using Self-Organizing Maps (SOMs) and Fuzzy Adaptive resonance Theory (ART) of unsupervised neural networks. In the experiment, we created an original facial expression dataset consisting of three facial expressions-happiness, anger, and sadness-obtained from 10 subjects during 7-20 weeks at one-week intervals. Results of creating FESCs in each subject show that the method can adequately display the dynamic diversity of facial expressions between subjects. Moreover, we used stress measurement sheets to obtain temporal changes of stress for analyzing psychological effects of the stress that subjects feel. We estimated stress levels of four grades using Support Vector Machines (SVMs). The mean estimation rates for all 10 subjects and for 5 subjects over more than 10 weeks were, respectively, 68.6 and 77.4%.
We propose a facial expression recognition method using several facial feature points associated with a majority of facial expressions. The facial feature points consist of the end points and centroids of eyebrows and eyes as well as several points around nose and mouth. We define a number of facial features based on the interrelation between some facial feature points including the distance between two facial feature points and the distribution of intensity values of the pixels in the region formed by three facial feature points. Since these facial features can be easily obtained from facial images with low computational cost, the proposed method performs efficient recognition. In addition, we propose a criterion to estimate the usefulness of facial features based on their variance ratio. By introducing feature extraction based on principal component analysis using only useful facial features, we develop an effective facial expression recognition method taking the tradeoff between recognition accuracy and recognition speed into consideration.
Conventionally speaking, a “smile” is considered an expression that characterizes the emotion of “happiness.” However, the word does not only refer to positive emotions such as happiness. It can also indicate negative emotions experienced in daily life (e.g., wry smile). This study categorizes expressions of five commonly used phrases with the word “smile” according to variation in expression and degree of positivity. As a result of a cluster analysis, χ2-test and ANOVA, three categories resulted: (1) somewhat negative smile with small variation, (2) smile with moderate variation and positivity, and (3) smile with large variation and positivity. In order to further study these characteristics, we analyzed the variations in individual facial expressions. The results showed that an increase in positivity was characterized by raised eyebrow ends, narrowing of the eyes, a wider mouth, and a smaller change in the corner of the mouth. Because variations in eyebrows and eyes are smaller than those in the mouth, there is a possibility that a smile is judged based on the change seen in the mouth. In order to analyze the psychological influence of the three types of smiles, personal impressions were also evaluated for each type of smile. We applied four factors that were extracted through factor analysis: favor ability, vitality, dominance, and femininity. The study revealed that the larger the variance, the higher the evaluation of vitality, dominance, and femininity, and the lower the evaluation of favor ability.
Image warping is an important visual effect tool in entertainment industry and other research field. And, it is often used also for correction of distortion caused by lens. To date many warping algorithms have been proposed and used such as mesh transformation, radial basis functions, thin plate splines and B-splines (free-form deformations). In particular, the free-form deformation based on B-spline approximation is a powerful and useful morphing algorithm and is proven to have the one-to-one property which can prevent the transformed image from folding back upon itself. In order to minimize the transformation error, a multi-level B-spline approximation has been proposed. Though a natural and smooth morphed image can be obtained by using the multi-level B-spline approximation, it takes a large computation cost. In this paper, we proposed an adaptive lattice partitioning method to reduce the large computation cost and a subregions transformation method for enhancement of transformation accuracy. The experimental results show that the proposed methods are more efficient and accurate than conventional multi-level B-spline method.
This study investigated impression factors related to the attractiveness evaluation of expressive faces, through rating experiments. In experiment1, the attractiveness rating on three types of expression form was conducted: neutral faces of neutral form not expressing any specific emotion, happy faces of positive form expressing a happy emotion, and sad faces of negative form expressing a sad emotion. As a result, positive form was more attractive than the neutral form, whereas the negative form was less attractive than the neutral form. In experiment2, the attractiveness rating and the impression rating by using the Semantic Differential method were conducted. Correlation analyses between attractiveness and the composite scores of impressions for each type of expression form showed that “Intellectual-beauty” impression was important for the attractiveness evaluation of neutral form and negative form, whereas “Mildness” impression as well as “Intellectual-beauty” impression was important for the attractiveness evaluation of happy face. These results suggest that the key impression factors related to the attractiveness evaluation were different for each type of expressive forms. There is a possibility that a common psychological criterion is used for attractiveness evaluation of neutral face and sad face, unlike with happy face.
The fields of psychology and engineering have amassed an extensive body of research concerning facial expression recognition. A great deal of this research employs a Facial Action Cording System (FACS) that facilitates facial expression recognition by tracking facial muscle movement. This approach requires a considerable amount of analysis to interpret facial muscle movement. We have attempted facial expression recognition by means of a simplified approach involving analysis of the sequential stages of a blink with respect to the resultant retinal outline. A subject presented a facial expression and the resultant sequential retinal outlines over the course of a blink were analyzed and stored in a database. This was done for seven different emotional expressions. When the same subject presented subsequent expressions for analysis the correct facial expression was identified at a high frequency by means of pattern matching with previously stored data. The fine musculature around the eye is recruited during facial expression formation and impacts the conformation of the eye lid over the course of a blink which in turn impacts the observable retinal outline. While our system is currently employed to recognize facial expressions of previously learned individuals there is a clear potential for further development.
There have been many studies about dynamics of neural fields. Especially, the neural field which allowed localized excitation areas provided base for the self-organizing map (SOM) algorithm. Here, we focus on a neural oscillatory field, and proposed a three-layered model of a neural oscillatory field which allows stable localized oscillatory excitation areas. Our comuputer simulation results show that the neural oscillatory field with two Mexican-hat-type connections keeps two or more than two localized oscillatory excitation areas stably around maximal points of an external input. In this case, the neural oscillatory field realizes in-phase phase-locking within each localized oscillatory excitation area, but maximizes the phase difference between different localized oscillatory excitation areas. This neural oscillatory field provides base for an oscillatory SOM algorithm, and will be useful to solve the binding problem with separated extraction of information.
Fuzzy c-means based classifier (FCMC) is a classifier based on clustering approaches. The classification accuracy on training sets can easily be improved by increasing the number of clusters. On the other hand, the accuracy on test sets (i.e., the generalization capability) is not necessarily improved by increasing the number of clusters. Especially when the number of training samples is relatively small, the classifier not only over-fits the data, but also obtains incorrect covariance matrices and cluster centers, since the number of samples in each cluster becomes small. Hence, the test set accuracy deteriorates. The performance of FCMC with 2 clusters in each class when the number of training samples is less than 1000 was already reported. This paper reports the scaling behavior of FCMC by testing with variously-sized training samples. The number of clusters of FCMC is increased up to 8. The number is not very large but FCMC in this paper has many parameters. LibSVM is one of the widely known state of the art tools of SVM classifier for large-sized data sets. The classification accuracy on test set, training time and testing time (i.e., detection time) of FCMC are compared with LibSVM by varying the number of training samples.