Face detection approach from video with the aid of KPCM and improved neural network classifier

In recent years, the detection of a human face from the video has become an interesting research topic due to the video surveillance and other security issues. Efficient face detection from the video has become an immense need as it can provide various identity measures in the field of defense and other security-related areas. In our proposed method we have developed an efficient method of face detection to index a particular face from different video shots. The proposed method can be divided into Different modules. In the first module, human face from the video is extracted using segmentation technique. In our proposed method, we have used Kernel-based Possibilistic C-Means for segmentation purpose. The second module in our method is the feature extraction process where shape, LBP, and some geometrical features are extracted. The various shape features like area, circularity, and eccentricity are extracted. Once the feature values are extracted we track the particular face using forward tracking process. After the tracking process, we employ the classification technique. The classifier we utilized here is the improved neural network where the weights factors are optimized using the modified cuckoo search algorithm. The performance is compared with some existing works in order to prove the efficiency of our proposed method.     

一种新的无监督前景目标检测方法

针对基于无监督特征提取的目标检测方法效率不高的问题,提出一种在无标记数据集中准确检测前景目标的方法.其基本出发点是:正确的特征聚类结果可以指导目标特征提取,同时准确提取的目标特征可以提高特征聚类的精度.该方法首先对无标记样本图像进行局部特征提取,然后根据最小化特征距离进行无监督特征聚类.将同一个聚类内的图像两两匹配,将特征匹配的重现程度作为特征权重,最后根据更新后的特征权重指导下一次迭代的特征聚类.多次迭代后同时得到聚类结果和前景目标.实验结果表明,该方法有效地提高Caltech-256数据集和Google车辆图像的检测精度.此外,针对目前绝大部分无监督目标检测方法不具备增量学习能力这一缺点,提出了增量学习方法实现,实验结果表明,增量学习方法有效地提高了计算速度.     

A view-invariant and anti-reflection algorithm for car body extraction and color classification

This study proposes an intelligent algorithm with tri-state architecture for real-time car body extraction and color classification. The algorithm is capable of managing both the difficulties of viewpoint and light reflection. Because the influence of light reflection is significantly different on bright, dark, and colored cars, three different strategies are designed for various color categories to acquire a more intact car body. A SARM (Separating and Re-Merging) algorithm is proposed to separate the car body and the background, and recover the entire car body more completely. A robust selection algorithm is also performed to determine the correct color category and car body. Then, the color type of the vehicle is decided only by the pixels in the extracted car body. The experimental results show that the tri-state method can extract almost 90% of car body pixels from a car image. Over 98% of car images are distinguished correctly in their categories, and the average accuracy of the 10-color-type classification is higher than 93%. Furthermore, the computation load of the proposed method is light; therefore it is applicable for real-time systems.     

基于谱域特征提取与线性回归分类的智能人脸识别算法

针对人脸识别中由于光线、表情变化和遮挡导致人脸图像变化的问题,提出了一种谱域特征提取与线性回归分类算法相结合的智能人脸识别方法。为了实现特征提取的目的,首先使用Viola-Jones算法从原始图像中提取初始人脸部分,并将其转换为120×120像素大小的灰度图像;然后提出了一种计算极坐标傅里叶变换（FFT）以获得预处理人脸图像主要幅度谱特征的新框架,进一步在预处理的图像上执行2D-DFT,并表示为1D P-FFT。特征值是1D P-FFT幅值中的最大值,提取的特征值用于构造表示人脸图像的符号对象。最后利用快速有效的线性回归分类算法实现分类。在AR和GT数据库上进行了各种实验,分别取得了97.51％和98.02％的准确率,与最近报道的一些人脸识别技术相比,提出的方法识别准确率更高。     

高分类率水下目标特征量提取方法研究

根据人眼分类双谱图时的特点,在双谱对称性所确定的三角形区域内提出了两种双谱幅值矩阵元素抽取方案,即沿平行于副对角线方向的抽取方案和沿列向量方向的抽取方案。对所抽取的元素,采用简单的求和或最大值操作进行特征提取,形成4种特征向量。利用支持向量机的一对一多分类方法进行目标分类,实验表明：在由双谱对称性确定的三角形区域内,采用沿平行于副对角线方向的元素幅值抽取方案,对所抽取元素幅值使用求和方法得到的特征向量具有非常高的正确分类率。由此方法获得的特征向量对于A、B、C三类水下目标辐射噪声的分类率达到了100%,得到的其他特征向量的平均分类正确率均稳定在95%以上。     

Micro‐Doppler classification of human movements using spectrogram spatial features and support vector machine

Accurate distinction of dynamic moving objects especially in the context of security surveillance attracts great attention of researchers and practitioners. In the same context, present study proposes an advancement in feature extraction method from the micro‐Doppler spectrogram with the application of spatial statistics for moving human subject classification which minimizes the spectrogram analysis. A novel approach of spatial feature extraction from whole image spectrogram, followed by support vector machine (SVM) classifiers algorithm for multiclass classification, has been proposed in the present study. The proposed method has been tested for prediction accuracy and validated by applying on a very close and important five distinct human activities (which usually arise at any security observation site) as reported in the available literature. The results obtained adopting the proposed approach exhibit high accuracy for multiclass classification; yielding cross‐validation accuracy of 96.7% while actual predication of testing data provides the accuracy of 93.33%. For the prediction of accurate data classes, the post‐processing of the spectrogram prior to feature definition has also been performed using spatial based methods to enhance micro‐Doppler signatures.     

梯度直方图卷积特征的胶囊网络在交通监控下的车型分类

为了充分利用图像信息以提高现有交通监控下车型分类的效果,在胶囊网络的基础上增加梯度直方图卷积（HOG-C）特征提取方法,提出HOG-C特征的胶囊网络模型——HOG-C CapsNet。首先,使用梯度统计特征提取层对图像中的梯度信息进行统计,构建方向梯度直方图（HOG）特征图;其次,使用卷积层提取出图像的颜色信息,把提取出的颜色信息与HOG特征图融合构成HOG-C特征图;最后,输入卷积层提取HOG-C特征图的抽象特征,并通过胶囊网络对提取的抽象特征进行具有三维空间特征表达的胶囊封装,使用动态路由算法实现车型分类。在BIT-Vehicle数据集上对该模型和其他相关模型进行的对比实验中,该模型得到98.17%的准确率、97.98%的平均精确率均值（MAP）、98.42%的平均召回率均值（MAR）和98.20%的综合评价指标。实验结果表明,该模型在交通监控下的车型分类上具有更好的效果。     

梯度直方图卷积特征的胶囊网络在交通监控下的车型分类

为了充分利用图像信息以提高现有交通监控下车型分类的效果，在胶囊网络的基础上增加梯度直方图卷积（HOG-C）特征提取方法，提出HOG-C特征的胶囊网络模型——HOG-C CapsNet。首先，使用梯度统计特征提取层对图像中的梯度信息进行统计，构建方向梯度直方图（HOG）特征图；其次，使用卷积层提取出图像的颜色信息，把提取出的颜色信息与HOG特征图融合构成HOG-C特征图；最后，输入卷积层提取HOG-C特征图的抽象特征，并通过胶囊网络对提取的抽象特征进行具有三维空间特征表达的胶囊封装，使用动态路由算法实现车型分类。在BIT-Vehicle数据集上对该模型和其他相关模型进行的对比实验中，该模型得到98.17%的准确率、97.98%的平均精确率均值（MAP）、98.42%的平均召回率均值（MAR）和98.20%的综合评价指标。实验结果表明，该模型在交通监控下的车型分类上具有更好的效果。     

Dynamic time warping based identification using gabor feature of Adaptive Motion Model for walking humans

In this paper, we propose a novel feature extraction method for the identification of humans. The main objective of our method is to identify each human being by extracting the Gabor feature based on the Adaptive Motion Model (AMM) for the motion of humans. In our method, the adaptive motion model, which can represent the temporal motion for each walking human is first made from the sequence images and, then, the Gabor features of the eight directions which can represent the spatial motion information for humans are extracted. The proposed feature extraction method can make a more accurate motion model by adjusting the weight between the previous and current model for each person. Moreover, our method has the advantage of allowing more information such as the Gabor features for the eight directions extracted from the AMM. Since the conventional method uses the face feature for each human being, it has disadvantages in the case of images of small size, while our method has better identification performance this case, because it only uses the spatio-temporal motion information. Finally, we identify each person by finding the minimum value of the extended dynamic time warping (DTW) for the eight Gabor features. The accuracy of the identification conducted using the proposed feature is better than that of the conventional method using the Gait Energy Image (GEI) and Face Image feature.     

Automatic environmental noise source classification model using fuzzy logic

Noise classification is very important nowadays. Fuzzy logic has been applied to many interesting problems in different areas including noise identification/recognition. With this study, we propose an automatic environmental noise source classifier based on fuzzy logic. The proposed classifier uses the feature parameters that are extracted using short-time auto-correlation function. Six commonly encountered non-stationary noise sources are chosen to recognize. These are subway, airport, inside car, inside train, restaurant, and rain. Classification accuracy of the proposed classifier ranged from 62% to 90% rates.     

道路线形特征自动分类方法

为了可持续性地更新道路信息数据库,利用车载全球卫星定位系统(GPS)产生的大量路径跟踪轨迹信息,快速捕捉道路信息变化,提出一个基于LVQ-Boosting的道路线形识别模型。该模型以学习向量量化（LVQ）为基础分类器,采用改进的Boosting算法进行网络集成,进一步提高LVQ的泛化能力,从而获得一个使用弱分类算法却具有强分类性能的分类器。该模型以GPS定位点坐标、速度和道路水平曲率为基本识别特征和输入变量,以道路线形特征为输出变量,实现自动识别道路线形特征,快速分组道路特征类型的目的。实验结果表明,该方法具有较高的道路线形的识别效率和精度。     

一种提取抖动视频中前景目标的新方法

提取视频中的前景目标信息是视频处理领域非常重要的问题,考虑到现实生活中会出现监控摄像头不可避免地会出现晃动或偏移情况,造成监控视频短暂抖动,此时背景图像灰度和纹理信息都会受到较大的影响,从而给后期进一步分析前景信息带来了巨大的困难。为了兼顾纹理特征提取和噪声抑制两方面的要求,针对抖动视频的前景提取问题,提出了一种有效的融合小波变换和在线混合高斯模型的方案。首先运用仿射变换在线逐帧校准,接着利用小波变换对图像去噪,并建立自适应模型迭代上述过程,最后利用在线混合高斯模型提取前景。实验结果表明,与同类方法相比,该算法无论针对单目标还是多目标视频均可以有效去除抖动,得到较好的前景目标提取效果,具有较高的准确性和鲁棒性。     

基于图像处理的变电站视频智能分析研究

针对电力系统在视频监视方面的无人值守需求,提出了一种基于图像处理技术的视频智能分析系统研究和实现方案。通过连通域分析、编码映射、Hough直线检测、轮廓检测、高斯混合前景检测、cvBlob目标跟踪等图像处理和特征提取技术对高清实时视频流进行分析,实现了几种设备状态识别和安防异常行为识别功能。实验结果验证了各分析功能的实时性和准确性,并对比部分功能在相关文献中的实验结论,以表明该系统方案的有效改进。     

轿车局部区域特征提取与车型识别

车型识别技术是智能交通管理系统的关键技术之一。目前车型识别多基于车牌及整体外形的识别技术。基于此提出从轿车图像提取局部信息,从而快速、准确的识别出轿车车型。主要研究轿车尾灯特征提取,并在此基础上进行轿车分类识别。采用变分水平集方法对轿车图像进行分割,获得描述尾灯区域特征的三个特征参数——宽高比、矩形度、分散度作为支持向量机的分类的输入特征向量,对31种不同车型进行分类与识别,准确率达到100%。此研究表明利用轿车的局部特征也能够达到轿车车型识别的效果,具有较高的应用价值。     

监控场景中的运动物体提取技术研究

针对监控场景中因存在遮挡而无法有效地提取出完整的运动序列这一问题,提出了一种将ViBe前景检测算法和改进后的粒子滤波跟踪算法相结合的跟踪提取方法。首先用ViBe来提取出场景中所有运动物体的前景轮廓;其次用粒子滤波来检测和跟踪目标物体;最后通过与目标物体的关联轮廓求交运算以及跟踪区域的反馈调节完成对目标物体运动帧序列的提取。当运动物体发生遮挡时,采用将跟踪区域内所检测到的前景轮廓重新加入到目标物体的关联轮廓中以保证后续可以继续用关联轮廓交集来提取。实验结果表明,该方法能够很好地保证提取的质量,并有效地解决了局部遮挡与全局遮挡情况下运动物体完整运动序列的提取。     

基于打印字符分析的打印文件检验研究*

研究了基于打印字符图像分析的计算机打印文件检验,以快速鉴别生成打印文件的源打印机.鉴别过程包括打印文件图像采集、图像预处理、特征提取和匹配.用自行设计的装置采集打印文件图像,经过打印字符提取和识别,利用距离变换和方向直方图方法对不同打印文件中的相同字符进行匹配,来进行打印文件的源打印机认定.在有40台激光打印机的数据库中测试,打印文件鉴别实验准确率达到89.51％,证明了本方法的有效性.     

L1 norm based pedestrian detection using video analytics technique

Pedestrian detection from images of the visible spectrum is a high relevant area of research given its potential impact in the design of pedestrian protection systems. In general, detection is made with two different phases, feature extraction and classification. Also, features for detection of pedestrian are already are available such as optimal feature model. But still required is an improvement in detection by reducing the execution time and false positive. The proposed model has three different phases, that is, background subtraction, feature extraction, and classification. In spite of giving entire information into feature extraction, the system gives only a useful information (foreground image) by twin background model. Then the foreground image moves to the feature extraction and classifies the pedestrian. For feature extraction, histogram of orientation gradient (HOG) L₁ normalization has been used. This will increase the detection accuracy and reduce the computation time of a process. In addition, false positive rate has been minimized.