This paper proposes the face recognition method using both the depth and infrared pictures. The conventional face recognition methods based on color picture recognize the human faces accurately, but are easily affected by the illumination and are vulnerable to the attempts to steal user's face information through fake face such as the photograph or the sculpture. On the other hand, the methods based on the depth or infrared picture are less affected by the illumination, and prevent an attempt to recognize a false face. This paper utilizes the depth picture to reduce the recognition time and the infrared picture to increase the recognition performance. In the face detection, this paper finds the nose of a person using the captured depth picture for reducing detection time, and it detects the region of the face. In the feature extraction, it extracts the feature pictures from the infrared picture by 3D Local Binary Pattern. In the face identification, this paper compares between the features of the captured face and features of faces that are pre-stored in the DB, and obtains the face similarity. If the similarity of the face is larger than the certain threshold, the face recognition succeeds. Simulation results show that the face recognition performance is good not only in the normal environment but also in the little illumination.
In this paper, we show reduction methods for search algorithms on graphs using quantum walks. By using a graph partitioning method called equitable partition for the the given graph, we determine “effective subspace” for the search algorithm to reduce the size of the problem. We introduce the equitable partition for quantum walk based search algorithms and show how to determine “effective subspace” and reduced operator.
By virtue of recent developments in machine learning techniques, higher-level information can now to be extracted from massive data. In this paper, we focus on extracting multiple semantic relations, using light-weight processing through the efficient low-dimensional expression of substrings in text data. We propose an approach to build features for relational classification consisted of only the low-dimensional vectors representing substrings between words called substring vectors . In addition, we investigate the relationship between the numbers of dimensions and the obtained accuracies when nonlinear classifiers are applied. The experimental results show that, with simple features and small computational cost, our approach using relatively low-dimensional representations achieves a sufficiently high accuracy that is better than most existing approaches.
Neural networks have a rich ability to learn complex representations and have achieved remarkable results in various tasks. However, they are prone to overfitting owing to the limited number of training samples and regularizing the learning process of neural networks is essential. In this paper, we propose a regularization method that estimates the parameters of a large convolutional neural network as probabilistic distributions using a hypernetwork, which generates the parameters of another network. Additionally, we perform model averaging to improve the network performance. Then, we apply the proposed method to a large model such as wide residual networks. The experimental results demonstrate that our method and its model averaging outperform the commonly used maximum a posteriori estimation with L2 regularization.
We numerically investigate the change in synchronization property in different frequency components associated with laser dynamics, which is termed dynamics-dependent synchronization, in two mutually-coupled semiconductor lasers. We introduce an optical amplifier to implement asymmetric feedback, and we change the feedback strength of one of the two coupled lasers to observe dynamics-dependent synchronization. In-phase synchronization is observed for the original signals, while anti-phase synchronization is found for the low-pass-filtered signals, in the presence of low-frequency fluctuation dropouts. We analyze dynamics-dependent synchronization by observing temporal changes in the short-term cross-correlation and the local optical frequency detuning.
Persistent homology computes the number and the widths of holes of a shape. The delay-coordinate embedding reconstructs an attractor from time series data. Recently the combination of persistent homology and attractor reconstruction is proposed for clustering time series data. We propose a criterion to determine the delay of attractor reconstruction for periodic and chaotically periodic signals. Our criterion chooses the delay that maximizes the hole width of reconstructed attractors. We compare it to the criterion with mutual information and that by Perea and Harer. It is found that the combination of our criterion with Perea and Harer's criterion works well.
Mechanism of two-tone suppression is studied using a coupled-oscillator model of the cochlea with feed-forward coupling. Local amplification of sound signals is modeled by using Stuart-Landau oscillators near the Hopf bifurcation, and transmission of sound signals is described as feed-forward coupling between the oscillators. Effect of suppressor signals on the response to probe signals is analyzed by numerical simulations. It is found that the effect of suppression is qualitatively different depending on relative frequency between probe and suppressor signals. By analyzing a simplified two-oscillator model, we explain the mechanism of the suppression, where configuration of the oscillators plays an essential role.
Hyperthermia is one of the modalities for cancer treatment, utilizing the difference of thermal sensitivity between tumor and normal tissue. In the treatment, tumor should be heated up to the therapeutic temperature between 42 and 45 °C. In Japan, large number of capacitive heating devices are employed for hyperthermic treatments. In addition, some effects, which do not need the heating more than 42 °C, such as mild hyperthermia have also been studied by use of the same heating device. Therefore, many papers, which describe effectiveness of these treatments, have been published. However, temperature distributions inside the patient body during the treatment have not been cleared. So, in this paper, numerical calculations of temperature distributions inside patient body during the capacitive heating have been described. As a result, surface of the patient body was easy to heat compared with deep region of the body. In addition, blood flow rate of a target had a large influence to its temperature.
Stochastic resonance (SR) is a noise-enhancement phenomenon that enables the detection of sub-threshold signals by adding noise and using nonlinear systems. This paper explores the applicability of SR in a BPSK receiver with sub-threshold signals. Although received signals are amplified as a result of the nonlinear behavior of the receiver, they are somewhat distorted. This results in the lower performance of SR receivers in comparison with linear receivers. Employing a parallel network of SR systems is expected to solve this problem. The present theoretical analysis demonstrates that in a certain noise intensity range, the output of the network can fully describe an input sub-threshold signal, and hence, the performance close to that of the linear receivers can be obtained. The effectiveness of the SR receiver was also demonstrated through a numerical example of the bit error rate (BER). However, achieving good BER performance requires an infinite number of arrayed SR systems, which is not realistic in practical systems. A design framework for an SR network with a finite number of elements and an appropriate noise intensity that can realize BER performance close to that in linear systems is also provided.