Toshi Minami

Identification of human faces based on isodensity maps

Abstract The feasibility of using isodensity lines for human face identification system is presented through experimental investigation. Instead of using feature points extracted from the faces, as is done in the conventional face matching systems, the technique presented uses gray level isodensity line maps of the faces. Only simple template matching is then required to match the individual isodensity lines. The preprocessing required, the properties of isodensity lines and some considerations for practical implementation are also discussed. The results show a 100% accuracy in matching same persons and a 100% accuracy in discriminating different persons (including persons with spectacles).

Human-face extraction using modified HSV color system and personal identification through facial image based on isodensity maps

In this paper, a modified HSV color system which is suitable for the extraction of the human face from general indoor backgrounds is proposed. Experimental results of a personal identification system based on isodensity maps is described. Computer simulation using 111 pictures with various backgrounds shows an 89% face extraction accuracy. For human face identification, experimental results using 3481 pairs of face images show a 99.9% accuracy.

Encoding of Line Drawings with a Multiple Grid Chain Code

The multiple grid(MG) chain code which uses four different square grids is proposed to encode line drawings. The main processes adopted in the code are: 1) a grid selection algorithm which allocates quantization points only to the vicinity of the course of a line drawing, 2) labeling rule on quantization points which makes the frequency of some codes larger than that of other codes, and 3) quantization points allocation-not to the corners, but to the sides of a square which makes the straight line segments larger without increasing quantization error. A performance comparison of various chain codes is made from the viewpoints of the encoding efficiency, naturalness of the encoded lines, and the rate distortion measure. Also, the superiority of the MG chain code to other codes is shown. At last, application of the MG chain code to the electronic blackboard system is explained.

Description and synthesis of facial expression based on isodensity maps

Methods of description and synthesis of facial expression based on isodensity maps are proposed in this paper. Previous methods have almost been based on the Facial Action Coding System (FACS), which describes facial expression using feature points extracted from facial parts such as eyes, nose, mouth, etc. However it is very difficult to extract such feature points, and to describe changes in density in variations such as wrinkles with feature points only. In contrast to these methods, isodensity maps are adopted in this paper as able to represent density variation of facial images and level planes. To describe facial expression, the physical state of a face must be described, and a tree structure, which is constructed based on the relationship among the loops of isodensity maps, is used for that purpose in this paper. Facial expression is synthesized by the characteristic points which are extracted from level planes. By using this method, changes of expression on the cheek and around the mouth are represented well between successive frames. The prospects of using these proposed methods for describing and synthesizing facial expression are very promising.

Identification of front-view facial images based on isodensity maps

In this paper, we describe a method for identifying facial images based on isodensity maps. An isodensity map reflects the 3D shape of the face in sets of closed loops that represent the connectivity of pixels with the same gray level and is a useful feature in facial descriptions. Identification begins by extracting the face region from the source image, then is followed by matching. The extraction of the face region is easily implemented using isodensity lines that represent facial contours. Two-stage matching composed of template matching of isodensity maps (global matching) and local density matching of isodensity lines (fine matching) was adopted. Matching tests were performed on images of faces which included people wearing eyeglasses, women wearing makeup, and unshaven men. A total of 3, 481 pairs (59 pairs of the same person and 3,422 pairs of different people) were used and 3,479 (99.9%, 57 pairs of the same person and 3,422 pairs of different people) were correctly identified. The experimental results demonstrated that this is an extremely effective method for personal identification and is sufficiently robust in practical use.

A character segmentation algorithm for mixed‐mode communication

In the realization of a mixed-mode communication, it is necessary before recognizing individual characters, to separate text and black and white figure regions and to extract efficiently the character region. Problems in such a procedure are the detection and correction of the inclination of a document, separation of contact characters, the merge of disconnected characters, and extraction of a letterhead. This paper describes the results of studies on such problems. The document considered is the English text image, containing binary figures and a letterhead. The basic idea is as follows. Connected regions are obtained by directional propagation and shrinking to merge figures. Then: (1) a thinning process is performed to detect the inclination angle of the input text; (2) the sizes of the connected regions and their relative locations are examined to extract the letterhead. Estimation of pitch is performed and statistical data about individual characters are used to separate contact characters or merge disconnected characters. The experiment was made for 10 and 12 pitch printed characters, and the correct extraction rate was 100 percent for individual characters, and 94.6 percent for a letterhead. Thus it was verified that the proposed method is useful in extracting the character region.

A restoration algorithm of fingerprint images

This paper discusses noise elimination and restoration of the ridge structure of fingerprints in preprocessing for automatic fingerprint identification. The basic idea is as follows. The object image is partitioned into 64 × 64 unit regions with 8 pixels as edge. After estimating the locations of the ridges and valleys the restoration is performed by gray-scale correction with different weights for each location. The processing is roughly divided as follows: (1) separation of background; (2) detection of ridge direction (quantized in 8 directions); (3) detection of ridge and valley locations; and (4) correction of gray-scale value. Processing (2) greatly affects the later accuracy of detection and correction and requires high accuracy. In this paper the direction of the ridges is estimated hierarchically both from the local and global views and then is corrected by applying stochastic relaxation. As a result of experiment for 28 fingers (75903 unit regions) 99.00% of the estimates were within direction error of 0 and ±1. Considering that true correspondence of feature points was achieved for more than 73% between the later processed result and the original image, the above value is quite satisfactory for identification, verifying the usefulness of the proposed method.

Description and matching of density variation for personal identification through facial images

The feasibility of using isodensity lines for a human face identification system is demonstrated through experi-mental investigation. Instead of using feature points extracted from the faces, as is done in the conventional matchingsystem, the technique presented uses gray level isodensity lines of the faces By comparing the shape of isodensitylines, people can be identified. The identification technique proposed has the following advantages over conventional

Model-based coding of facial images based on facial muscle motion through isodensity maps

A model-based coding system has come under serious consideration for the next generation of image coding schemes, aimed at greater efficiency in TV telephone and TV conference systems. In this model-based coding system, the senders model image is transmitted and stored at the receiving side before the start of the conversation. During the conversation, feature points are extracted from the facial image of the sender and are transmitted to the receiver. The facial expression of the sender facial is reconstructed from the feature points received and a wireframed model constructed at the receiving side. However, the conventional methods have the following problems: (1) Extreme changes of the gray level, such as in wrinkles caused by change of expression, cannot be reconstructed at the receiving side. (2) Extraction of stable feature points from facial images with irregular features such as spectacles or facial hair is very difficult. To cope with the first problem, a new algorithm based on isodensity lines which can represent detailed changes in expression by density correction has already been proposed and good results obtained. As for the second problem, we propose in this paper a new algorithm to reconstruct facial images by transmitting other feature points extracted from isodensity maps.

Reconstruction of curved surface using isodensity maps based on stereo vision system

The passive stereo vision systems is one of the most general techniques of obtaining 3D information from objects. A common problem in stereo systems is finding the corresponding points in the two images of a scene. To cope with this problem, edge-based stereo methods have been considered useful. For curved surfaces, however, edge-based stereo cannot be applied because of the difficulty of extracting edges from curved surfaces. In this paper, we propose a new stereo algorithm based on isodensity lines extracted from a curved surface. Isodensity lines are extracted by tracing pixels having the same density. The proposed algorithm consists of two processes. The first is the process of finding corresponding isodensity lines between two images. The second process is finding one-to-one pixel correspondence along corresponding isodensity lines, for fine matching. With this algorithm, objects with curved surfaces can be reconstructed--even objects having no edges--and poorly corresponding points can be corrected easily. Though some problems are still to be investigated thoroughly, computer simulation results show good accuracy. From the experimental results, the use of this method for the 3D reconstruction of objects with curved surfaces can be considered a serious possibility.