An efficient image segmentation is important for computer vision in any attempt to analyze or interpret an image automatically. The automatic segmentation based on color in a natural image is very hard. So, semi-automatic segmentation methods incorporating user interactions are becoming more and more popular. Our approach makes possible that the image can be separated in many regions starting from a mean shift over-segmentation followed by the user interactions. After the over-segmentation, the user selects the regions of interest and makes the interactive operations. In order to avoid the fatigue in the user and to decrease the user interactions, three operations are made semi-automatically using the polygonal representation of the regions. The polygons are constructed using a new points connection algorithm marking branch points and removing overlapped connections. The proposed interactive image segmentation method is compared with well-known related algorithms in terms of the accuracy and efficiency. Although the accuracy of the proposed segmentation result depends on the user interactions, our method increases the object and boundary accuracy in 8% and 20% respectively in comparison with graph cuts methods using less time. Compared with an algorithm based on the results of mean-shift segmentation, proposed approach needs more user interactions but the segmentation results are more accurate according to Jaccard index and visual information, mainly in the boundary part.

View full abstract

A highly accurate extraction of facial areas from natural background for the purpose of the recognition of facial expressions is presented. To improve the recognition accuracy of facial expressions, the facial area must be extracted precisely before extracting facial parts such as eyes, mouth, and so on. Only using moving information and the color information associated with facial area, the facial area is extracted without using facial parts, in this system. By adopting the three phase estimating algorithm of facial skin color areas, the estimation accuracy is extremely improved. In order to extract moving area, an inter frame subtraction method with variable frame interval was adopted and the adequate frame interval was set. From the computer simulation using 34278 video frames, 33366 images (97.34%) were correctly extracted. From the experimental results, we show the prospects of using this method.

View full abstract

In this paper, we propose a region-based damage detection method using post-earthquake aerial images and a digital map. In our method, we first define the damage level in each region using image features, such as edges and textures. Then, we use the probabilistic relaxation method by transiting the level with the region similarity, and we obtain the damage probability. In this computation, we increase the detection accuracy by using two-step processing of non-damaged region detection and damaged region detection. We can obtain the final results both on the building segments and the road segments. Our method does not generate the results from two alternatives, such as damaged or not. The final judgement is committed to users, accordingly our method has advantages of automatic detection and accuracy of visual judgement.

View full abstract

The mean-shift tracking algorithm uses the tracker that scans locally according to the gradient of color features so that targets can be estimated faster. A method of determination of the histogram resolution used for constructing the weight distribution is proposed. The method uses the variance ratio of the target and the background color histograms. Actually, some experimental results showed the effectiveness of the proposal method. And the correlation between the background region area size and the tracking accuracy of proposed method is shown by some experimental results.

View full abstract

Recently, services using facial information such as age and gender are provided widely. This paper proposes a mean to estimate age and gender at once in facial images by using novel features. First, a facial image is normalized by using Active Appearance Model (AAM), then appearance parameters are used as global feature. Next, Gabor magnitude pictures are obtained by convolving the normalized image with Gabor filters, followed by encoding with Local Directional Pattern (LDP) operator which enhances information. Then, the maps are divided into several blocks, and histograms are extracted from each block. The histograms are concatenated to a vector which is adopted as local feature. Then PCA is used to reduce the dimensions. Next, the global and local features are combined to one vector. Finally, age is estimated by using SVM and SVR, and also gender is classified by using SVM. In the experiment, the proposed age estimation method improves 0.7 or more on Mean Absolute Error (MAE), and the result of gender classification on average is 89.35% that is best performance as compared with the conventional methods. As a result, the experiment demonstrates that the proposed algorithm is an effective method, compared to the other similar methods.

View full abstract

This paper proposes a new approach for simply and precisely mapping a real texture of a two-dimensional (2D) object surface in a camera image onto surfaces of a model object and a virtual object in mixed reality space by using the 2D cross-ratio technique. The object surfaces treated here are planar quadrilaterals with four vertices. The 2D surface texture mapping is based only on the cross-ratios among the vertices. The main advantage of the proposed approach is that the three-dimensional perspective projection transformation inevitably required in the conventional methods is not needed at all, which means the principle and procedure for the mapping are very simple. A real-image experiment demonstrates the effectiveness of our proposed approach.

View full abstract