压力状态下人脸高光谱图像波段选择方法研究

发布时间：2018-02-02 02:06

本文关键词： 高光谱波段选择压力识别多头绒泡菌算法线性预测禁忌搜索　出处：《西南大学》2017年硕士论文　论文类型：学位论文

【摘要】：针对生理信号非接触识别压力的研究是情感计算的一项重要内容,而基于高光谱技术提取的人脸血氧含量可用于压力的非接触识别。高光谱图像在包含丰富信息的同时,其庞大的数据量、众多的波段、相邻波段极强的相关性等也给压力识别的实时处理带来了很大的困难,解决这个问题的关键就是减少高光谱图像的波段数量,波段选择是最为常用的波段降维方式,故本文在基于高光谱技术提取的人脸血氧含量用于压力识别的基础上,针对高光谱图像的波段选择算法进行研究。主要工作如下:(1)提出了一种改进的基于多头绒泡菌模型的高光谱图像波段选择算法。多头绒泡菌模型能够解决最短路径问题,该问题与高光谱波段选择中寻找最小冗余波段组合的问题相似。本文将多头绒泡菌算法应用到高光谱波段选择中,通过绒泡菌网络自组织、自优化的特性对波段进行选择。此外,为了实现波段的自动选择,提出了基于互信息的非均匀子空间划分方式,同时采用了最佳指数倒数作为适应函数参与波段选择。实验结果表明使用非均匀划分子空间方式比传统的均匀划分方式的效果更好,采用最佳指数倒数这种适应函数选择的波段效果比采用传统的相关性好,使用多头绒泡菌算法选择的波段组合的压力识别效果接近于全部波段,证明了算法的有效性。(2)提出结合线性预测与禁忌搜索的高光谱图像波段选择算法。结合压力状态下的人脸高光谱数据,将线性预测获取的较优组合作为禁忌搜索的初始解,以压力的分类准确率作为适应函数,利用禁忌搜索算法优化波段组合使之具有更高的压力分类准确率。实验结果表明结合禁忌搜索算法选择的波段对压力的识别更精确,证明了算法的有效性。(3)设计心理压力诱发实验,获取了25名普通大学生的面部高光谱图像数据,包括心理压力和平静状态。此外,对采集的数据进行预处理,提出了一种基于血氧直方图的端元自动提取方法,将该方法产生的样本光谱曲线数据用于波段选择;采用了最小噪声分离法去除高光谱数据中的噪声,实验结果表明去噪后的数据比未去噪数据产生的面部血氧图更清晰,更利于波段选择后的压力识别。本文利用采集的高光谱数据对提出的两种算法做了验证。通过验证分析表明,本文提出的两种波段算法在保持压力识别准确率的情况下,能够大量的减少波段数量,对基于高光谱技术的压力识别来说是有效的波段选择方法。
[Abstract]:The research on non-contact recognition pressure of physiological signals is an important part of emotional calculation. However, the extraction of blood oxygen content based on hyperspectral technology can be used in the non-contact recognition of pressure. Hyperspectral images contain rich information, its huge amount of data, numerous bands. The extremely strong correlation of adjacent bands also brings great difficulties to the real-time processing of pressure recognition. The key to solve this problem is to reduce the number of bands of hyperspectral images. Band selection is the most commonly used wave band dimension reduction method, so this paper based on the extraction of facial blood oxygen content based on hyperspectral technology for pressure recognition. The band selection algorithm of hyperspectral images is studied. The main work is as follows: 1). An improved band selection algorithm for hyperspectral image based on multi-headed actinomycetes model is proposed, which can solve the shortest path problem. This problem is similar to the problem of finding the least redundant band combination in the hyperspectral band selection. In this paper, we apply the multi-headed actinomycetes algorithm to the hyperspectral band selection, and self-organize through the chorionic bacteria network. In addition, in order to realize the automatic band selection, a non-uniform subspace partition method based on mutual information is proposed. At the same time, the optimal exponent reciprocal is used as the adaptive function to participate in the band selection. The experimental results show that the non-uniform partition subspace method is more effective than the traditional uniform partition method. The effect of the band selection using the best index reciprocal fitness function is better than that of the traditional correlation, and the pressure recognition effect of the band combination selected by the multi-headed actinomycetes algorithm is close to that of the whole band. Proved the effectiveness of the algorithm. 2) proposed a combination of linear prediction and Tabu search hyperspectral image band selection algorithm. Combined with the pressure of the hyperspectral face data. The optimal combination obtained by linear prediction is regarded as the initial solution of Tabu search, and the classification accuracy of pressure is taken as the fitness function. The Tabu search algorithm is used to optimize the combination of bands to achieve higher accuracy of pressure classification. The experimental results show that the band selected by Tabu search algorithm is more accurate to identify the pressure. Proved the validity of the algorithm. 3) designed the psychological stress induced experiment, obtained the facial hyperspectral image data of 25 ordinary college students, including psychological stress and calm state. Based on the preprocessing of the collected data, a method based on the blood oxygen histogram is proposed to extract the endcomponents automatically. The sample spectral curve data generated by this method are used to select the band. The minimum noise separation method is used to remove the noise from the hyperspectral data. The experimental results show that the data after denoising is more clear than the facial blood oxygen map produced by the non-denoising data. In this paper, we use the collected hyperspectral data to verify the proposed two algorithms. The two band algorithms proposed in this paper can greatly reduce the number of bands while maintaining the accuracy of pressure recognition, which is an effective band selection method for pressure recognition based on hyperspectral technology.
【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】