A novel method of using the spiking neural networks (SNNs) and the electroencephalograph (EEG) processing techniques to recognize emotion states is proposed in this paper. Three algorithms including discrete wavelet transform (DWT), variance and fast Fourier transform (FFT) are employed to extract the EEG signals, which are further taken by the SNN for the emotion classification. Two datasets, i.e., DEAP and SEED, are used to validate the proposed method. For the former dataset, the emotional states include arousal, valence, dominance and liking where each state is denoted as either high or low status. For the latter dataset, the emotional states are divided into three categories (negative, positive and neutral). Experimental results show that by using the variance data processing technique and SNN, the emotion states of arousal, valence, dominance and liking can be classified with accuracies of 74%, 78%, 80% and 86.27% for the DEAP dataset, and an overall accuracy is 96.67% for the SEED dataset, which outperform the FFT and DWT processing methods. In the meantime, this work achieves a better emotion classification performance than the benchmarking approaches, and also demonstrates the advantages of using SNN for the emotion state classifications.

We present a detailed study on the RSS-based location techniques in wireless sensor networks (WSN). There are two aspects in this paper. On the one hand, the accurate RSSI received from nodes is the premise of accurate location. Firstly, the distribution trend of RSSI is analyzed in this experiment and determined the loss model of signal propagation by processing experimental data. Secondly, in order to determine the distance between receiving nodes and sending nodes, Gaussian fitting is used to process specific RSSI at different distance. Moreover, the piecewise linear interpolation is introduced to calculate the distance of any RSSI. On the other hand, firstly, the RSSI vector similarity degree (R-VSD) is used to choose anchor nodes. Secondly, we designed a new localization algorithm which is based on the quadrilateral location unit by using more accurate RSSI and range. Particularly, there are two localization mechanisms in our study. In addition, the generalized inverse is introduced to solve the coordinates of nodes. At last, location error of the new algorithm is about 17.6% by simulation experiment.

In this paper, mass spectral substance detection methods are proposed, which employ long short-term memory (LSTM) recurrent neural networks to classify the mass spectrometry data and can accurately detect chemical substances. As the LSTM has the excellent understanding ability for the historical information and classification capability for the time series data, a high detection rate is obtained for the dataset which was collected by a time-of-flight proton-transfer mass spectrometer. In addition, the differential operation is used as the pre-processing method to determine the start time points of the detections which significantly improve the accuracy performance by 123%. The feature selection algorithm of Relief is also used in this paper to select the most significant channels for the mass spectrometer. It can reduce the computing resource cost, and the results show that the network size is reduced by 28% and the training speed is improved by 35%. By using these two pre-processing methods, the LSTM-based substance detection system can achieve the tradeoff between high detection rate and low computing resource consumption, which is beneficial to the devices with constraint computing resources such as low-cost embedded hardware systems. INDEX TERMS Mass spectral substance detections, long short-term memory networks, chemometrics.

No abstract

No abstract

Light field photography captures 4D radiance information of a scene. Digital refocusing and digital correction of aberrations could be done after the photograph is taken. However, capturing 4D light field is costly and tradeoffs between different image quality metrics should be made and evaluated. This paper explores the effects of light field photography on image quality by quantitatively evaluating some basic criteria for an imaging system. A simulation approach was first developed by ray-tracing a designed light field camera. A standard testing chart followed by ISO 12233 was provided as the input scene. A sequence of light field raw images were acquired and processed by light field rendering methods afterwards. Through-focus visual resolution and MTF were calculated and analyzed. As a comparison, the same tests were taken for the same main lens system as the results of conventional photography. An experimental light field system was built up and its performance was tested. This work helps better understanding the pros and cons of light field photography in contrast with conventional imaging methods and perceiving the way to optimize the joint digital-optical design of the system.

No abstract

This paper proposes three methods to improve the learning algorithm for spiking neural networks (SNNs). The aim is to improve learning performance in SNNs where neurons are allowed to fire multiple times. The performance is analyzed based on the convergence rate, the concussion condition in the training period and the error between actual output and desired output. The exclusive-or (XOR) and Wisconsin breast cancer (WBC) classification tasks are employed to validate the proposed optimized methods. Experimental results demonstrate that compared to original learning algorithm, all three methods have less iterations, higher accuracy, and more stable in the training period.