Hirotsugu Kashimura

Classification of human actions using face and hands detection

In this paper, we describe a novel classification technique that separates video scenes, like office work tasks, into several scenes according to each task. Even if the difference of as a whole image frame by frame in each task is small, the difference of workers movement is quite big due to the position of face and hands according to each task. In addition, the worker has the tendency to turn his/her face to look at the particular objects of individual tasks like PC, a document, and so on. Then, we decide to separate tasks based on face position, face angle in depth, and hand positions. For comparison of frames in a video, we use the Maharanobis distance to measure the difference of multivariate data that consist of face coordinates, face angle in depth, and coordinates of both hands. For the separation of tasks by the Maharanobis distance of an each frame, we use the hierarchical clustering method to classify frames in a video according to each task. For the robust detection of both hands, we use color-based method that searches hand areas using face color. Although the color of hands changes corresponding to the lighting conditions, the color of hands should be very similar to that of the face in an office. Therefore, even when the lighting condition changes, our color-based hand detection method does not need any adjustment to the change. We apply this classification technique to separate office work video into individual sets of task scenes. As a result, our technique shows better task separation performance than the histogram-based boundary detection technique.

Size-Independent Image Segmentation by Hierarchical Clustering and Its Application for Face Detection

In this paper, we introduce a technique to detect a target object quickly. Our idea is based on onservation on the clusters into which an image is divided by hierarchical k-means clustering with space feature and color feature. This clustering method has the advantage of extracting the region of an object with some varied size. We insist that our idea should lead to detect a target object quickly, because it is not necessary to search the locations containing no targets. First, we evaluate our clustering method and second, we demonstrate that our method is effective on an object detection by applying to our face detection system. We show that the detection time can be reduced by 24%.

Scaling, rotation, and translation invariant image recognition using competing multiple subspaces

We propose a tolerant object recognition system under a combination of various transformations of an object image. The system realizes invariant recognition by re-normalizing the image with multiple units each of which is assigned to the individual transformation. The re-normalization process is an iterative procedure in which only the most accurate unit re-normalizes the images every iteration. To implement the re-normalization units, we utilize a kernel-based non-linear subspace model. In the model, projection of the image to the subspace represents the amount of transformation in the manner of the population coding. In addition, the accuracy of the representation can be known as distance between the image and the subspace. The system is applied to face detection from snapshots to show significant robustness under scaling, rotation, and translation.

a-Si:H TFT-driven high-gray-scale contact image sensor with a ground-mesh-type multiplex circuit

An A4 page width and 300 dot/inch hydrogenated amorphous silicon thin film transistor (a- Si:H TFT) driven contact image sensor which can read more than 128 gray levels has been developed. Crosstalk due to the coupling between data lines in the multiplex circuit has prevented high gray scale reading. In order to eliminate crosstalk, a sensor with a new multiplex structure has been developed with a ground mesh shield layer inserted at the crossover points between each data line. The ground mesh shield pattern was designed to optimize the gray scale reproduction ratio R. With this sensor, R is more than 0.992 for a single bit, thus achieving 128 levels of gray. This design was compared to the performance of two other sensors, one without a ground mesh shield, the other using a data line meander pattern. This technology is also applicable to higher performance image sensors with greater than 400 dot/inch resolution.

Face recognition using surface features in XYI space

We propose a face recognition algorithm that utilizes novel surface features in (x, y, I(x,y)) space. A face image is considered as a surface in XYI space, and the surface is segmented into a definite number of regions by using a Gaussian mixture model. Parameters of each Gaussian distribution are determined by maximizing the log-likelihood function, and are stored as features of each individual face image. In the recognition process, the log-likelihood is used as a similarity measure between a test image and the stored features. The face recognition performance of our algorithm is evaluated with the FERET database. Our algorithm achieves an identification rate of 95.4% and equal error rate of 1.4%, which are superior to other algorithms based on eigenface features and Gabor wavelet features.