非相似度保持投影

由主成分分析(PCA)可知任何一幅人脸图像都可以通过一组特征脸的线性加权来重构,PCA是最小均方误差意义下图像的最优表示,但是传统的PCA最终只通过比较加权系数的欧氏距离来进行识别,没有考虑残差。因此,提出非相似尺度的概念,将两个样本同时投影到相同向量上,在确定它们关系时既考虑投影系数,也考虑重构所产生的残差。两者的投影系数和残差相差越大,说明这两个样本越不相似。和保局投影(LPP)有所不同,非相似度保持投影算法不必预先设定近邻个数,它是利用非相似度的概念,创建非相似度散布矩阵,最终通过最大化目标函数获取最优子空间。在AR库和Feret库上的实验结果证明了该方法的有效性。     

航空发动机传感器故障鲁棒检测方法

研究发动机传感器故障准确检测问题,现代航空发动机数字电子控制系统对传感器的可靠性要求日益提高。针对航空发动机结构复杂,又工作在高温和高压下,常规采用的传感器故障检测方法的准确性易受到建模误差与外界扰动的影响,造成漏报或误报。为了提高检测精度,提出建立航空发动机数控系统传感器未知输入故障模型,采用特征结构配置的方法,通过配置闭环系统左特征向量实现故障检测残差对不确定性因素的干扰解耦,降低扰动对故障诊断结果的影响。用某型涡扇发动机数控系统传感器故障数字仿真试验表明,所设计的方法对范数有界的不确定量可以实现干扰解耦,抑制干扰对故障检测的影响,改善检测算法的鲁棒性,提高检测结果的准确性,同时满足在线运算的实时性要求。提高了航空发动机的可靠性,保证了安全飞行。     

基于特征互补图像的人脸识别仿真研究

研究人脸识别精度问题。由于人脸图像中存在大量干扰信息的缺点,而造成了人脸识别正确率下降,为了解决上述问题,提出了一种基于特征互补图像快速特征融合算法。算法通过对人脸图像的位平面切片图像分析,采用位平面图像分解法,通过各种合成策略构造多幅样本图像。并突出高位平面图像,采用两种加权策略将每一幅人脸图像样本都生成"特征互补图像"。然后,直接用图像的二维典型相关分析(2DCCA)法对两种特征互补图像进行特征抽取。最后通过在ORL国际标准人脸库上进行的实验,结果表明,高位平面图像的典型相关鉴别特征提高了正确识别率,并且因为摒弃了原始人脸图像的大部分干扰信息所以具有更强的鲁棒性。     

一种谱分解降维的模糊有监督局部保持投影策略

提出一种谱分解降维的模糊有监督局部保持投影策略。首先针对监督局部保持投影SLPP存在过学习和不能较好地保持图像空间的差异信息等问题,通过最小化局部离散度和最大化差异离散度准则提取投影方向,找到一种线性鉴别分析的等价形式。其次,通过采用模糊k近邻(FKNN)方法得到相应的样本分布隶属度信息,同时考虑到离群样本对整个分类结果的不利影响,提出一种模糊化方法,根据样本的隶属度对样本分布矩阵重定义所做的贡献,将每个样本的隶属度融入到SLPP特征抽取的过程中,从而得到完整有效的模糊样本特征向量集,有效解决了小样本问题的特征抽取问题。第三,提出一种谱分解的矩阵分析方法,在SLPP投影准则下,对散布矩阵实现降维。在ORL和NUST603人脸库上的实验结果验证了该方法的有效性。     

基于图像仿射不变特征点的零水印算法

本文提出了一种基于图像提升小波变换和仿射不变特征点的零水印算法。首先对图像进行3级提升小波变换,然后在低频部分,利用Harris-Affine算子提取出图像的仿射不变特征点和仿射协变特征区域,利用这些特征区域来构造零水印信息。实验证明,该算法不仅对常规信号处理攻击和简单的几何攻击具有很强的鲁棒性,对于剪切、纵横比改变、行列去除,局部扭曲等较复杂的几何攻击也具有较强的抵抗能力。     

Reciprocal bi-directional generative adversarial network for cross-modal pedestrian re-identification

To improve the accuracy of cross-modal pedestrian re-identification,a reciprocal bi-directional generative adversarial network-based method is proposed.First,we build two generative adversarial networks to generate cross-modal heterogeneous images.Second,an associated loss is designed to pull close the distribution of features in latent space during the image translation between visible and infrared images so as to help the networks generate fake heterogeneous images that have high similarity with the real images.Finally,by concatenating the original and generated heterogeneous pedestrian images into the discriminative feature extraction network,images from different modalities can be unified into a common modality,thus suppressing the cross-modal gap.Representation learning and metric learning are utilized to achieve more discriminative pedestrian features.Comparative experiments are conducted on SYSU-MM01 and RegDB datasets to analyze the accuracy with different loss functions.Compared with other state-of-the-art cross-modal pedestrian re-identification methods,the proposed method achieves a higher accuracy and stronger robustness.     

Method for discrimination of false targets in multistation radar systems based on the deep neural network

For the existing jamming discrimination methods for multistation radar systems,only the single feature of target echo space correlation is utilized as the metric,which leads to insufficient comprehensiveness of feature extraction,so that effectiveness and universality are insufficient for the discrimination algorithm.In this paper,an identification method in multistatic radar systems based on the deep neural network is proposed.This method combines the characteristics of multistatic radar systems cooperative detection technology,which has many available resources and strong scheduling ability in space,time and frequency domain,with the strong model learning and feature representation ability in the process of information processing on the deep neural network,so that it can effectively apply to the field of anti-deception jamming.Full use is made of unknown information about echo data to obtain more multi-dimensional,more comprehensive,more complete and deeper feature differences besides correlation,so as to achieve a better jamming discrimination effect.Simulation results show that the proposed method can effectively reduce the influence of noise and pulse number on the jamming discrimination performance.At the same time,the limitation of the target echo correlation coefficient on anti-jamming technology under nonideal conditions is alleviated,which broadens the boundary conditions of the application process.     

Fast network intrusion detection system using adaptive binning feature selection

Aiming at the problems of the low detection rate of traditional intrusion detection systems and the long training and detection time of intrusion detection systems based on deep learning,an adaptive binning feature selection algorithm using the information gain is proposed,which is combined with LightGBM to design a fast network intrusion detection system.First,the original data set is preprocessed to standardize the data;then the redundant features and noise in the original data are removed through the adaptive binning feature selection algorithm,and the original high-dimensional data are reduced to the low-dimensional data,thereby improving the accuracy of the system and reducing the training and detection time;finally,LightGBM is used for model training on the training set selected by the characteristics to train an intrusion detection system that can detect attack traffic.Through verification on the NSL-KDD data set,the proposed feature selection algorithm only takes 27.35 seconds in feature selection,which is 96.68% lower than that by the traditional algorithm.The designed intrusion detection system has an accuracy rate of 93.32% on the test set,and its training time is low.Compared with the existing network intrusion detection system,the accuracy rate of the proposed system is higher,and its model training speed is faster.     

一种面向弱纹理图像的特征点描述子

现有的大多数特征点提取算法适用于处理纹理丰富的图像,而对于弱纹理图像则无法提取有效的特征点. 对此,提出了多邻域结构张量特征(MNSTF)算法. 基于一系列固定的邻域和图像结构张量,通过表达局部图像的结构和纹理信息,解决了弱纹理和无纹理场景下特征点提取和匹配等相关问题;同时,通过计算邻域之间的相对方向,实现了MNSTF算法特征描述子的旋转不变性. 实验结果表明,MNSTF算法在经过旋转的弱纹理图像测试集上的特征点匹配准确率达到了99.9%以上,验证了其良好的适用性、旋转不变性和鲁棒性.     

基于改进MobileNet v2的垃圾图像分类算法

针对现有的垃圾图像分类模型实时性能差和分类精度低的问题,提出基于改进MobileNet v2的垃圾图像分类方法,构建以MobileNet v2为核心的轻量级特征提取网络. 通过调整宽度因子降低模型的参数量;在模型中嵌入通道和空间注意力模块,增强网络对特征的细化能力;设计多尺度特征融合结构,增强网络对尺度的适应性;利用迁移学习的方式优化模型参数,进一步提高模型精度. 实验结果表明,算法在自建数据集上的平均准确率为94.6%,分别高于MobileNet v2、VGG16、GoogleNet、ResNet50、ResNet101模型2.0%、3.4%、3.2%、2.3%、1.2%;所提算法在2种公共图像分类数据集CIFAR-100和tiny-ImageNet中均取得不错表现;模型参数量仅为0.83 M,体积约为基础模型的2/5,在边缘设备JETSON TX2上的单次推理耗时68 ms,实现了推理速度和预测准确率的提升.