针对单一特征在描述方言间差异性方面存在不足和传统高斯混合通用背景模型(Gaussian Mixture Model-Universal Background Model,GMM-UBM)在训练时存在混叠的问题,在将经验模式分解引入到特征提取的基础上,融合多特征形成高维特征.进一步,通过多次训练GMM-UBM,筛选出最具区分性的方言模型以提升各方言模型间的区分性.方言种类识别对比实验结果表明,基于融合特征与改进GMM-UBM的方法优于传统方法.

Aiming at the problem that the current defocus blurred region detection methods are easy to misjudge homogeneous-sharp regions, and can′t locate the edge accurately enough, a defocus blurred region detection method based on LBP and saliency is proposed. Firstly, LBP feature and SLIC algorithm are used to obtain SLBP blur map, DRFI saliency detection algorithm is used to obtain DRFI saliency map. Secondly, the trimap is obtained using SLBP blur map and DRFI saliency map, and then KNN matting algorithm is used to obtain a blur map. Finally, the blur map is refined with the help of morphological operations and smoothing filter. The experimental results on the public blur dataset show that the method can effectively detect homogeneous-sharp regions, retain the edge details of an image, and perform well in detection accuracy and recall.

经典的i-vector的提取方法利用方言特征在通用背景模型(Universal Background Model,UBM)的统计差异来构建全局差异空间,对方言语种的区分能力较弱.为此,提出了一种基于改进的i-vector的提取算法,利用方言特征在方言相关的高斯混合模型(Gaussian Mixture Model,GMM)上的统计差异来构建全局差异空间,提升i-vector对方言语种的区分能力.首先基于方言相关GMM分别构建全局差异空间;其次拼接各空间中提取到的i-vector并进行主成分分析(Principal Component Analysis,PCA)降维,得到改进的i-vector;最后采用高斯概率线性判别分析(Gaussian Probabilistic Linear Discriminant Analysis,GPLDA)模型进行建模和打分.实验表明,所提算法较经典i-vector算法能更有效地提升对方言语种的识别性能.

为实现更具真实感的表情图像合成,探讨一种基于自注意力机制和谱归一化的生成式对抗网络(GAN)表情合成方法.通过在生成器中引入2层自注意力模块,使生成器能够在局部建立丰富的上下文关系,输出更加真实的表情细节;同时,在鉴别器中引入谱归一化,使鉴别器的训练更加稳定.实验结果表明,该模型在主观视觉和FID图像评价指标上均优于其他典型算法,图像质量和表情细节有明显提高.

针对非局部均值算法在合成孔径雷达(Synthetic Aperture Radar,SAR)图像的边缘和纹理区域去噪效果较模糊的问题,探讨改进了非局部均值算法的权重部分,把高斯加权的空间距离与非局部均值算法的权重系数相结合,构造新的权重系数,提出了一种基于改进非局部均值算法的SAR图像去噪算法.首先,使用对数变换处理图像;其次,进行高斯滤波;再次,利用了所提出的算法将高斯滤波结果去噪;最后,采用指数变换方法处理去噪结果.去噪结果表明,所提算法在抑制散斑噪声和保持图像结构信息方面有较好的性能.

图像匹配是计算视觉领域的基本问题,多谱段图像匹配一直是图像匹配中的一个难点,被广泛应用于土地覆盖、变化检测分析、图像融合中.在利用FFT实现相位相关的基础上,针对其存在频谱泄露和边缘效应的问题,提出一种加窗预处理的相位相关法,对合成孔径雷达图像和光学图像进行匹配,并通过核回归方法生成二维窗函数.实验结果表明基于加窗的相位相关对多谱段图像匹配有效.

Dual-channel contrast prior (Dual-CP) simulates contrast using the difference between the bright channel and the dark channel of an image, and it achieves good results in the blind restoration of blurred images. However, in practical applications, the values of the bright channel and the dark channel of an image are not distributed on 1 and 0 as theoretically researched. This paper proposes a blind image restoration algorithm that combines Dual-CP, L-0 regularization strength, and gradient prior, wherein an effective optimization algorithm is derived using semi-quadratic splitting method to solve the nonconvex L-0 minimization problem. Experiments demonstrate that the proposed method has better intuitive description recovery capabilities, and on the benchmark dataset presented by Levin et al., Kohler et al., and Lai et al., the average peak signal-to-noise ratio increased by 2. 1051 dB, 1. 1273 dB, and 0. 4491 dB, respectively, and the average structural similarity increased by 0. 1302, 0. 0599, and 0. 0158, respectively.

针对面部动力谱(FDM)特征易受光照影响、对运动信息描述不准确的缺陷,提出基于FDM特征与时空局部二值模式积分投影(STLBP-IP)特征相结合的微表情识别方法.将FDM特征与STLBP-IP特征相结合,在弥补FDM对运动信息描述不足的同时对人脸面部信息进行补充描述以提升识别率.使用支持向量机进行分类,在SMIC和CASMEⅡ微表情数据库上进行实验.实验结果表明,该算法识别率有所提高,分别达到57.14％和64.59％.

In recent years, continuous growing interests have been seen in bringing artificial intelligence capabilities to mobile devices. However, the related work still faces several issues, such as constrained computation and memory resources, power drain, and thermal limitation. To develop deep learning (DL) algorithms on mobile devices, we need to understand their behaviors. In this article, we explore the architectural behaviors of some mainstream DL frameworks on mobile devices by performing a comprehensive characterization of performance, accuracy, energy efficiency, and thermal behaviors. We experimentally choose four model compression methods to perform on networks and in addition, analyze the related impact on the nodes amount, memory, execution time, model size, inference time, energy consumption, and thermal distribution. With insights into DL-based mobile application characteristics, we hope to guide the design of future smartphone platforms for lower energy consumption.

城市交通流具有非线性、相关性、周期性等特点,其建模和准确预测难度较大.为提高预测准确率,在雨天、非雨天和工作日、非工作日四种影响因素下搭建基于长短期记忆(LSTM)网络模型的双向长短期记忆(BiLSTM)网络模型用于提取不同因素下的数据周期特性,利用实测数据进行交通流预测分析实验.结果表明,与不考虑影响因素即随机环境相比,在考虑影响因素下BiLSTM模型预测的效果稍好;同时,在四个影响因素下,与传统RNN模型、LSTM模型进行比较,BiLSTM模型预测效果均稍好.

复制粘贴篡改是最常见的图像篡改方式之一,常用检测算法中,SIFT算法是基于点特征的检测方法,对小的篡改区域检测性能较差.针对这一问题,本文探讨通过增加MB-LBP纹理特征提取更多的特征匹配对和提升检测定位能力的算法.实验结果表明,小区域的篡改检测定位是可行的.

针对传统GrabCut算法需要人工初始化而引起图像分割效率低的问题,结合目标轮廓增强技术,提出一种自动GrabCut算法.首先对图像进行谱残差计算,以获取目标轮廓增强的视觉显著图;其次,对显著图进行预分割并通过快速连通区域分析进行前景估计并获取掩膜,将获取的掩膜代替人工交互初始化GrabCut算法,最后实现自动分割.实验结果表明,该方法克服了手动的缺点,并在处理前后景颜色相似的图像时,分割结果要优于传统方法.

针对基于常规特征提取的交互行为识别算法在大样本数据集上难以提取有效特征、浅层卷积神经网络(CNN)分类错误率较高等问题,研究一种基于HSV颜色空间图像增强和深层CNN的改进算法.首先提取视频数据的RGB帧,扩充数据集,在HSV颜色空间下图像增强,然后微调CNN模型的最后一个全连接层和soft-max层,最后将增强后的图像输入到VGG16网络模型中进行学习和识别分类.在BIT-Interaction数据集上做交互行为识别实验,结果验证了所探讨算法在交互行为识别上的有效性.

To address the problems of human-computer interaction, poor separation, low efficiency, and difficulty in seed selection in the traditional image foreground-background separation algorithms, we propose an automatic image-foreground-background separation algorithm based on the texture features extracted in the Lab color space. First, we segment the image into blocks and convert it into a CIE-Lab color space established by the international commission on illumination (CIE). Then, we extract the color and texture features of each image block and select seeds. Finally, we use a region growing algorithm for image separation and region merging to reduce over separation. The experimental results show that the proposed algorithm is superior to the traditional algorithm in terms of the separation results, processing time, and algorithmic complexity.

为降低晶圆缺陷对半导体制造的影响,设计一种基于机器视觉的晶圆表面缺陷在线自动检测技术.首先,针对晶圆中单个晶元可能出现缺角或者遮挡,设计一种基于轮廓匹配的晶元定位方法;其中选取轮廓完整的良品图,经过图像拉伸、中值滤波、边缘检测、形态学处理,得到外边缘轮廓和内部圆点轮廓,建立外边缘轮廓模板,搜索晶圆图像金字塔进行模板匹配.其次,针对晶元背景是有一定规律排列的几何图案,缺陷可能与背景几何图案相似的问题,利用仿射变换原理,设计一种几何图案轮廓仿射变换与分区域检测方法;在此基础上,裁剪几何图案内部区域和几何图案外部区域,分别进行阈值分割和形态学处理,提取缺陷.再次,将缺陷做并集,得到总的缺陷,并进行标记,完成缺陷检测.实验结果表明,所提检测算法能有效解决晶元缺角或严重遮挡无法匹配定位和缺陷与背景几何图案相似的问题,且能有效检测出各种缺陷,单个小晶元的检测速度约430 ms,算法效果好,检测速度快,符合工业要求.

针对用于人脸微表情识别的面部动力谱特征(FDM)提取方法中迭代算法抽取时空立方体主方向时时间复杂度高的问题,提出基于主元分析(PCA)的改进算法.首先将数据中心平移到原点,其次估计相关矩阵,最后估计其特征向量和特征值,取最大特征值对应的特征向量方向作为时空立方体的主方向.采用Oulu大学SMIC微表情数据库中的微表情片段作为实验样本,选择支持向量机(SVM)作为分类器,对改进算法和原算法进行人脸微表情识别对比实验.结果表明,两种算法识别率相近,但改进算法在计算时间上远短于原算法.可见,改进算法在准确找出时空立方体主方向的同时,能大大降低原算法的计算复杂度.

By introducing the idea of optimal feature representation in kernel-based nonlinear representor (KNR), the traditional support vector machine (SVM) is improved in this manuscript to a new version called KNR-SVM. Firstly, the k-means unsupervised method is adopted to cluster the input patterns to form k-centers. Then, a mapping function, a point of a reproducing kernel Hilbert space (RKHS), is estimated based on the k-centers and KNR method, and the input patterns are mapped explicitly into a higher dimensional feature space, using the estimated function. Finally, the traditional SVM is taken as a classifier to classify the mapped patterns. Experimental results on handwritten digital recognition show that the proposed KNRSVM outperforms the traditional SVM.

An improved method was proposed for the forgery detection and tampering localization of double JPEG image compression based on the feature of first digit of DCT coefficients. First, a to-be-checked JPEG image was divided into overlapping blocks size 64 × 64 and then in every 8 × 8 block of each image block the first digit (1~9) features of AC-DCT coefficients in the first nine positions were extracted. Secondly, PCA dimension and compression were adopted to process the extracted first digit features for compact features and then Support Vector Machine ( SVM) classifier was used to judge whether the corresponding image block had been re-compressed. Fi-nally, the test results of double compression were used to locate the tampered area of JPEG image. Experimental results show that in comparison with representative algorithms, the improved algorithm produces better tampering localization and is robust against operations such as rotation, scaling and feathering.

Recently, mobile applications have gradually become performance- and resource- intensive, which results in a massive battery power drain and high surface temperature, and further degrades the user experience. Thus, high power consumption and surface over-heating have been considered as a severe challenge to smartphone design. In this paper, we propose DTEHR, a mobile Dynamic Thermal Energy Harvesting Reusing framework to tackle this challenge. The approach is sustainable in that it generates energy using dynamic Thermoelectric Generators (TEGs). The generated energy not only powers Thermoelectric Coolers (TECs) for cooling down hot-spots, but also recharges micro-supercapacitors (MSCs) for extended smartphone usage. To analyze thermal characteristics and evaluate DTEHR across real-world applications, we build MPPTAT (Multi-comPonent Power and Thermal Analysis Tool), a power and thermal analyzing tool for Android. The result shows that DTEHR reduces the temperature differences between hot areas and cold areas up to 15.4°C (internal) and 7°C (surface). With TEC-based hot-spots cooling, DTEHR reduces the temperature of the surface and internal hot-spots by an average of 8° and 12.8mW respectively. With dynamic TEGs, DTEHR generates 2.7-15mW power, more than hundreds of times of power that TECs need to cool down hot-spots. Thus, extra-generated power can be stored into MSCs to prolong battery life.

An improved forgery detection and tampering localization method is proposed for double JPEG compression images based on first digit feature of discrete cosine transform (DCT) coefficients and kernel-based nonlinear representor (KNR). First, a to-be-checked JPEG image is divided into overlapping blocks of size 64×64, and the first digit (1~9) features of alternating current (AC) DCT coefficients at the first nine positions of every 8×8 blocks are obtained in each image block, followed by principal component analysis (PCA) transform to form compact features. And then the KNR classifier is used to judge whether the corresponding image block has been re-compressed. Finally, the test results of double compression are used to locate the tampered area of JPEG image. Experimental results show that in comparison with representative algorithms, the improved algorithm achieves better results, and is robust to operations such as rotation, resizing and feathering. Moreover, KNR classifier outperforms classical SVM classifier in recognition effect and efficiency.