结合高斯滤波与MASK的G-MASK人脸对抗攻击

深度神经网络的快速发展使其在计算机视觉和自然语言处理等领域取得较大成功,但是对抗攻击会导致神经网络的表现性能降低,对各类系统的安全保密性造成严重威胁。现有黑盒攻击方法在人脸识别中性能表现较差,攻击成功率较低且生成对抗样本迁移性不高。为此,提出一种结合高斯滤波与掩码的对抗攻击方法G-MASK。利用Grad-CAM输出的热力图确定对抗样本的掩码区域,使其只在掩码区域施加扰动,提高黑盒攻击成功率,采用扰动集成方法提高黑盒迁移能力,增强黑盒攻击鲁棒性,对生成的扰动进行高斯平滑处理,降低集成模型之间干扰噪声的差异,提高图像质量且增强扰动掩蔽性。实验结果表明,针对不同的人脸识别模型,G-MASK方法在保证白盒攻击成功率较高的条件下能够显著提升黑盒攻击效果,并具有更优的掩蔽性,经过模型扰动集成的对抗样本白盒攻击成功率均提高至98.5%以上,黑盒攻击成功率最高达到75.9%,与快速梯度符号法(FGSM)、迭代快速梯度符号法(I-FGSM)和动量迭代梯度符号法(MI-FGSM)相比分别平均提升12.1、10.6和8.2个百分点,充分验证了该方法的有效性。

Color Image Enhancement by an Integral Mask-filtering Approach Employing Nonlinear Transfer Function

We demonstrate a brand-new method to sharpen a color image by using an integral mask-filtering technique. The derivatives between the target pixel and its neighbors are transferred by the cubic root function instead of the traditional linear one. The obtained final image has clearer fine characteristics along with much less overshooting.

Spatial mask filtering algorithm for partial discharge pulse extraction of large generators

A spatial mask filter algorithm （SMF） for partial discharge （PD） pulse extraction is proposed in this paper. In this algorithm, firstly, a ＇Teager＇ operator is used to strengthen wavelet coefficient local energy; then direct multiplication of coefficients at two adjacent scales is used to detect singularity points of the signal and to obtain scale based spatial mask filter; finally, an ＇ AND＇ logic operator is used in different filters to obtain the last spatial mask filter. By multiplication of wavelet coefficients with the final mask filter and wavelet reconstruction process, partial discharge pulses are extracted. The results of digital simulation and practical experiment show that this method is superior to traditional wavelet shrinkage method （TWS）. This algorithm not only can increase the signal to noise ratio （SNR）, but also can preserve the encrgy and pulse amplitude.

基于Mask R⁃CNN与SG滤波的手势识别关键点特征提取方法

手势识别是人机交互的重要手段。为了精确识别手势并摒除光照等环境干扰,同时减除由于手部高维运动造成的关键点剧烈抖动的问题,提出一种基于基于蒙版区域的卷积神经网络(Mask Region⁃based convolutional neural network,Mask R⁃CNN)与多项式平滑算法(Savitzky⁃Golay,SG)的手势关键点提取方法。该方法首先对输入的红绿蓝(RGB)三通道图像进行特征提取与区域分割,获得手部的实例分割与掩码。然后利用ROIAling及功能性网络进行目标匹配,标记出22个关键点(21个骨骼点+1个背景点)。将标记后结果送入SG滤波器进行数据平滑,并进行骨骼点的重新标定。从而得到稳定的手势提取特征。对模型进行对比实验,结果表明,该方法能够最大程度摒除环境干扰,并精准提取关键点。与传统基于轮廓分割的手势关键点提取相比,模型的鲁棒性大大提高,识别精度达到93.48%。

基于Mask R-CNN的铸件X射线DR图像缺陷检测研究

针对传统的铸件缺陷检测不能对缺陷进行分类分级等问题,提出了一种基于Mask R-CNN的铸件X射线DR图像缺陷检测算法。首先对原始图像进行预处理,采用引导滤波进行图像平滑,平滑图像与原图像进行差分得到差分图像,将差分图像与平滑图像相加运算使图像增强,再利用Labelme进行图像标注,形成训练数据集。送入Mask R-CNN深度学习网络,通过特征提取网络生成建议区域,分类、回归网络生成边界框和掩码,经多次参数调节后得到训练网络模型,最后测试数据集。实验数据结果表明,气泡1~5级的检测率分别为:66.7%,71.4%,77.4%,88.9%,87.5%;疏松1~5级检测率为:62.5%,72.2%,77.1%,83.3%,81.1%。检测结果证明应用Mask R-CNN结合引导滤波增强方法的缺陷检测方法可以较好的实现对铸件X射线DR图像的缺陷检测的分级分类,为工业铸件缺陷检测提供了应用深度学习方法的解决方案。

一种改进的基于Mask原理的DMC影像匀光算法

DMC影像普遍存在亮度不平衡的现象,在使用前一般都需要进行匀光处理。文章利用DMC影像中间区域受曝光不均影响较小的特点,在通过高斯滤波获得背景影像后,改进Mask匀光法直接将原影像和背景影像作相减,并将背景影像中间区域的亮度值作为基准对原影像进行匀光,提出了针对DMC影像的匀光处理流程。实验验证,匀光效果理想。

改进的自适应Mask匀光算法

Mask匀光算法在光学遥感影像色彩一致性处理中广泛应用。针对传统算法存在滤波器参数确定自适应性差、影像存在反差不一致现象、影像细节信息易损失等缺陷,提出改进算法。该算法结合影像信息与最优化准则自适应确定滤波器参数,对背景进行动态压缩以减小细节信息丢失,自适应调整拉伸参数,解决反差分布不一致现象。对比试验显示,改进算法处理的影像在光照一致性、影像清晰度、反差一致性上均取得较为理想的匀光效果。

Matched Filter Based Spectrum Sensing When Primary User Has Multiple Power Levels

In this paper,we investigate the matched filter based spectrum sensing in a more reasonable cognitive radio(CR) scenario when the primary user(PU) has more than one transmit power levels,as regulated in most standards,i.e.,IEEE 802.11 Series,GSM,LTE,LTE-A,etc.This new multiple primary transmit power(MPTP) scenario is specialized by two different targets:detecting the presence of PU and identifying the power level.Compared to the traditional binary sensing where only the presence of PU is checked,SU may attain more information about the primary network(making CR more "intelligent") and design the subsequent optimization strategy.The key technology is the multiple hypothesis testing as opposed to the traditional binary hypothesis testing.We discuss two situations under whether the channel phase is known or not,and we derive the closed form solutions for decision regions and several performance metrics,from which some interesting phenomenons are observed and the related discussions are presented.Numerical examples are provided to corroborate the proposed studies.

Continuously Variable Bandwidth Sharp FIR Filters with Low Complexity

In software defined radio (SDR), sharp filters of different bandwidth are required to fine tune the desired channel. This requires different computational resources and large number of filter coefficients. This paper proposes a continuously variable bandwidth sharp finite impulse response (FIR) filter with low distortion and low complexity. For this, a fixed length FIR filter is used with two arbitrary sampling rate converters. This system can be used for both the continuous increase as well as decrease of the effective bandwidth of a filter. The low complexity and sharpness are achieved by using the frequency-response masking (FRM) approach for the design of the fixed length FIR filter. The sharp transition width leads to maximum rejection to channel interference in SDR.

Design of Sharp 2D Multiplier-Less Circularly Symmetric FIR Filter Using Harmony Search Algorithm and Frequency Transformation

In this paper, we present a novel and efficient method for the design of a sharp, two dimensional (2D) wideband, circularly symmetric, FIR filter. First of all, a sharp one dimensional (1D) infinite precision FIR filter is designed using the Frequency Response Masking (FRM) technique. This filter is converted into a multiplier-less filter by representing it in the Canonic Signed Digit (CSD) space. The design of the FRM filter in the CSD space calls for the use of a discrete optimization technique. To this end, a new optimization approach is proposed using a modified Harmony Search Algorithm (HSA). HSA is modified in such a way that, in every exploitation and exploration phase, the candidate solutions turns out to be integers. The 1D FRM multiplier-less filter, is in turn transformed to the 2D equivalent using the recently proposed multiplier-less transformations namely, T1 and T2. These transformations are successful in generating circular contours even for wideband filters. Since multipliers are the most power consuming elements in a 2D filter, the multiplier-less realization calls for reduced power consumption as well as computation time. Significant reduction in the computational complexity and computation time are the highlights of our proposed design technique. Besides, the proposed discrete optimization using modified HSA can be used to solve optimization problems in other engineering disciplines, where the search space consists of integers.

基于信息增强和掩码损失的红外与可见光图像融合方法

针对低光场景下红外与可见光融合图像中存在的细节弱化和边缘模糊等问题,提出一种基于信息增强和掩码损失的图像融合方法。首先,采用引导滤波增强可见光图像的纹理细节和红外图像的边缘梯度;其次,构建双分支特征提取网络提取不同模态图像的特征信息,并设计交互增强模块以渐进交互的方式集成不同特征分支的互补信息,增强特征的细节表示;然后,在融合阶段设计注意力引导模块从空间和通道维度上关注特征信息,提升网络对关键特征的感知能力;最后,提出一种掩码损失以指导融合网络有针对性地保留源图像信息,提升融合质量。为验证所提方法的融合性能,在MSRS、TNO和LLVIP公开数据集上与9种主流的融合算法进行实验对比。结果表明,所提方法在定性和定量评估上均优于其它对比算法,生成的融合图像具有丰富的纹理细节、清晰的显著性目标和良好的视觉感知。

Design of Computationally Efficient Sharp FIR Filter Utilizing Modified Multistage FRM Technique for Wireless Communications Systems

Modern wireless communications gadgets demand multi-standard communications facilities with least overlap between different input radio channels. A sharp digital filter of extremely narrow transition-width with lower stop band ripples offers alias-free switching among the preferred frequency bands. A computationally competent low pass filter (LPF) structure based on the multistage frequency response masking (FRM) approach is proposed for the design of sharp finite impulse response (FIR) filters which are suitable for wireless communications applications. In comparison of basic FRM with other existing multistage FRM structures, the proposed structure has a narrow transition bandwidth and higher stop band attenuation with significant reduction in terms of the number of computational steps. A design example is incorporated to demonstrate the efficiency of the proposed approach. Simulation results establish the improvement of the proposed scheme over other recently published design strategies.

Method Investigation for Fabricating Micro-Optical Elements by Use of Half Tone Masks

Several approaches to fabricate micro optical elements by use of half tone masks are studied and compared. It is shown that half tone masks employed in filtering image systems can obtain gray patterns with considerably high precision, but it is hard to operate from the viewpoint of operation. The method using contacting lithography technology or laser ablation can be easily operated with the cost of reducing fabrication precision and the trouble of choosing appropriate materials. For all of these methods, the coding of half tone masks with corrections for the nonlinear characteristics of coding, imaging and photoresist is recommended.

基于改进Gabor算法的遮挡人脸智能识别方法

为提高有遮挡人脸的识别精度,提出基于改进Gabor算法的遮挡人脸智能识别方法。首先,对人脸图像动态范围压缩,并选择反锐化掩模滤波算法展开图像增强处理;其次,利用Gabor滤波器对信息保留较完整、亮度较高的半边脸进行特征提取;最后将提取到的Gabor特征输入到极限学习机中完成遮挡人脸的智能识别。实验结果表明,所提方法对处理遮挡人脸图像具有良好的效果,且其对人脸图像识别具有精准度高、识别时间短等优点。

基于Mask R-CNN倾斜影像筛选的建筑物三维模型高效重建方法

为了解决村镇区域建筑实景三维建模效率低下的问题,提出一种基于影像筛选的三维模型重建方法.首先,利用网络开源的建筑检测数据集训练Mask R-CNN神经网络模型;其次,利用训练好的神经网络模型筛选包含建筑的倾斜影像;最后,使用下视影像和筛选的倾斜影像对研究区进行三维重建.本研究使用Mask R-CNN神经网络模型从9775张无人机影像中自动筛选出7451张包含建筑的影像.实验结果表明,基于影像筛选的增量式三维重建法比常规方法节省了49.4%的数据处理时间,其空三重投影均方根误差和密集重建结果与常规方法一致.

信息熵-低通滤波联合掩模云层干扰去除方法

为了降低海洋光学遥感图像中云区对海面表面纹理分析的影响,本文开展了信息熵-低通滤波融合掩模的云层干扰去除研究。首先,分析了现有遥感图像去云算法的基本原理及其局限性。在此基础上,提出一种基于信息熵-低通滤波联合掩模的云层干扰去除方法。其中包括对遥感图像的改进矩匹配的去条带预处理、局部信息熵滤波,以及联合低频滤波作为遥感图像中各像元的校正参数。该算法具有复杂度低、处理速度快的特点。实验结果表明:与现有算法相比,本文提出的算法可在低计算复杂度的前提下,大幅增强各区域的纹理细节信息,其平均信息熵可达到7.8以上,对比度可达到60以上,平均梯度可达到200以上;图像细节方面,本文算法能够在不引入伪边缘、非均匀性的前提下,充分展现受云区影响的海表纹理细节,进一步满足高保真度的遥感应用需求。

过滤式消防自救呼吸器的改进对性能影响研究

针对过滤式消防自救呼吸器在长时间贮存和穿戴时分别存在的易变形和不舒适的问题,提出了一种半面罩改进、进气口改进和密封形式改进的优化方法,用以提高可贮存性、人体穿戴体验和避免对消防救援过程产生负面影响。探讨了目前呼吸器的包装、进气口设计、半面罩材料存在的缺陷,依此提出了消防呼吸器设计的具体改进方案,并通过漏气系数、吸气阻力、呼气阻力和一氧化碳防护性能试验对改进方案进行可行性验证。经试验证明,改进后产品相较于原有产品性能变化显著:其中呼吸区漏气系数降低超过80%,吸气阻力降低超过30%,呼气阻力降低超过60%,一氧化碳透过质量浓度降低45%。经优化的包装形式和采用新型的半面罩可有效解决现有存储和穿戴不适问题,更加契合面部形态、缩短穿戴时间,同时增加使用的安全性。

基于掩码的改进滤波算法

工业机器人图像处理中的噪声去除是一个重要且具有挑战性的问题。传统的图像滤波算法在处理工业机器人图像中的噪声时存在局限性,无法同时处理不同类型和强度的噪声。基于此提出了一种基于掩码的工业机器人图像滤波技术,通过掩码分割和多种滤波算法的组合,能够有效地去除工业机器人图像中的混合噪声,并提高图像处理的准确性和稳定性。实验证明,该方法在去除椒盐噪声和高斯噪声方面取得了良好的效果,为工业机器人图像处理领域提供了一种可行的方案。

基于同相轴优化追踪的多次波匹配衰减方法

提出基于同相轴优化追踪的多次波匹配衰减方法,首先利用SRME预测得到的多次波记录生成多次波速度谱,然后应用同相轴追踪技术提取多次波同相轴旅行时信息,再通过同相轴优化追踪实现对多次波预测误差的校正,最后应用短时窗的FK滤波对多次波予以压制。该方法仅根据多次波同相轴旅行时信息进行匹配衰减,降低对多次波预测精度的要求,并且能够显著提高多次波压制的精度。理论模型实验与野外资料处理实验结果均表明,该方法的多次波剔除效果明显优于常规的多道维纳滤波方法与抛物线拉冬域蒙版滤波法。

交流接触器开关电弧的图像采集与处理

综合利用计算机视觉技术、光电子技术、CCD(Charge Coupled Device)高速成像技术和计算机图像处理技术研制了一种集图像采集、传输、存储、处理于一体的新型电弧图像高速采集系统,利用该系统可拍摄到低压电器开关电弧从燃弧到灭弧的一系列图像。为有效地抑制噪声、提高信噪比,本文首先应用自适应中值滤波算法去除随机噪声,然后采用拉普拉斯掩模锐化算法突出图像的边缘和细节,使处理后的电弧运动图像边缘清晰可辨,从而适于对电弧直径变化进行分析计算。通过对处理后的电弧从燃弧到灭弧的一系列图像的分析理解,归纳出低压交流接触器分断电弧的燃弧过程的特点,且发现所得的电弧运动图像特征与电弧的大量仿真结果相一致。