There is a rising demand for intelligent robot cameras capable of shooting dynamic pictures like those taken by professional broadcasting studio cameramen. To develop thes cameras, we devised a system for analysing the actual movements of broadcasting-studio cameramen. We conducted experiments by using programs shot in our studio.
From the experiment, we found that cameramen routinely move their eyes in a range of about ±20 degrees horizontally and±15 degrees vertically before they start shooting, and they often look through the view-finder at the script to focus. We also found that the panning-speed curve is asymmetrical : quick acceleration is followed by slow deceleration. The deceleration time is 60% longer than the acceleration time, and the maximum acceleration during panning is approximately 200 deg/sec².

View full abstract

Blinking was studied as a physiological index to objectively evaluate the psychological influence of 3D-HDTV images on viewers. Variance of blinking frequency, which indicates the localization of blinks occurring for each cut quantitatively, was used as the method of evaluation. When compared to the variance of blinking frequency while viewing 2D-HDTV images, the results showed that the variance for 3D-HDTV was larger than that for 2D-HDTV. This suggests that the occurence of blinking depends more on the content of each cut for 3D-HDTV than it does for 2D-HDTV. To examine why, cuts that caused frequent blinking are discussed from the viewpoint of interest or concentration. The number of blinks per minute, measured over several cuts, is also investigated. When the viewing time is about 18 minutes, the number of blinks was greater while viewing 3D images than whil eviewing 2D images for one subject tested. For four other subjects tested, however, there was no significant difference in the number of blinks.

View full abstract

Miniature eye movement is characterized by it's complex waveform, small amplitude and unconsciousness. Involuntary movement like this may be affected by various psychological and physical states. From a view point of a fractal dimension analysis, the characteristics of miniature eye movements were investigated for estimating our sensitivity of feeling. The correlation dimension was used as a fractal dimension for quantizing complexity of miniature eye movements drift components. Validity of sufficient total calculation data (i. e. 18, 000 points for 12 embedding dimensions using a bandpass filter) were confirmed. And also differences between experimental data and noise of measurement system were simulated using pseudo white noise. And 1/f power spectrum of experimental data was confirmed using Fast Fourier Transform. This suggested that miniature eye movement may be a biological chaos. The applications of this analysis may be not only sensitivity of feeling estimation but also a medical application as a diagnosis of dementia.

View full abstract

A neural net approach is proposed to classify the eye shape with the respect to the three dominant features obtained from the subjective evaluation standard. The neural network is composed of eighteen physical features of the eye used as input data and the three features in the subjective tests used as output data. From the experiments, it can be concluded that these features combined can best represent various subjective impressions of human eyes. It can also be shown that unknown data can be classified by using k-neighbor interpolation training, where neighboring data are chosen based on the criterion of absolute standard of learning data.

View full abstract

A method for recognizing facial expressions from time-sequential images by using Hidden Markov Models (HMM) is proposed. HMM has the advantage that it can process time-sequential infomation. Moreover we can expect the HMM to make generalizations from the training data because of its learning procedure. Each image of a facial expression is transformed into an image feature vector. Each element of the feature vector consists of the average power from a distinct frequency band obtained by applying the Wavelet transformation to the image. The sequence is converted into a symbol sequence by using a new category-separated vector quantization. The codebook is constructed by appending codewords selected from other categories to each category to reduce the probability of wrong symbolization for similar facial expressions. To recognize an observed sequence, the HMM that best matches the sequence is chosen, and the category of the HMM is the recognized expression. Experiments for recognizing 4 expressions result in a promising recognition rate of 93.7%.

View full abstract

This article introduces a newly devised orienting task and discusses its utility in face processing research. In a pilot study investigating feature saliency, subjects were presented with a series of line-drawn faces and were asked to select 10 of their favorite faces, or 10 faces that were most similar to a particular target. The line-drawn faces were randomly selected from a pool of nearly 60, 000 different faces, which were drawn by choosing one of 3 possible values for each of 10 facial features (i. e., 3¹⁰ possible faces). Assuming that the subjects would select faces with consistent values for important features, the relative importance of facial features may be assessed by comparing the amount of variance “tolerated” by the subjects for each feature. We found that, for example, in similar face retrieval, the variance was significantly smaller for eyebrow tilt, eye shape, and face shape than for other features such as eyebrow position, eye position, or nose length. The utility of the present orienting task is discussed in terms of what these experimental results reveal about feature saliency.

View full abstract

An objective evaluation measure AWSNR (Advanced Weighted Signal to Noise Ratio) which compensates coding noise by considering entropy in every 16×16 pixels sub-block is proposed. In order to clarify the relationship between MOS (Mean Opinion Score), SNR and entropy, a subjective evaluation test using twelve kinds of textures with ten different noise levels is carried out. After the noise compensation factor is experimentally derived from the entropy-SNR characteristics when MOS=4, 200 images with five kinds of coding noise or with random noise are evaluated subjectively and objectively. The results show an improved correlation coefficient with the MOS-AWSNR of 0.91, compared to 0.66 with the MOS-SNR.

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract