基于欧氏空间相似度的云模型协同过滤算法

传统的基于余弦相似度度量的云模型协同过滤推荐算法未考虑特征向量的长度和维度,忽略了三个重要数字特征云期望、熵和超熵的关系,如各数字特征具有不同的性质和权重,导致特征丢失、区分度过小的问题。针对这些问题,提出了一种采用标准化的多维欧几里德相似度计算方法,通过将三个数字特征映射为三维空间的点,计算经指数函数标准化的欧几里德相似度,生成更合理的用户k近邻集,最终产生推荐。实验结果表明,该相似度计算方法能够为云特征向量提供更显著的区分度,并在一定程度上提高了推荐质量。     

Face Orientation Normalization Using Eye Positions

Despite the fact that progress in face recognition algorithms over the last decades has been made, changing lighting conditions and different face orientation still remain as a challenging problem. A standard face recognition system identifies the person by comparing the input picture against pictures of all faces in a database and finding the best match. Usually face matching is carried out in two steps： during the first step detection of a face is done by finding exact position of it in a complex background （various lightning condition）, and in the second step face identification is performed using gathered databases. In reality detected faces can appear in different position and they can be rotated, so these disturbances reduce quality of the recognition algorithms dramatically. In this paper to increase the identification accuracy we propose original geometric normalization of the face, based on extracted facial feature position such as eyes. For the eyes localization lbllowing methods has been used： color based method, mean eye template and SVM （Support Vector Machine） technique. Experimental investigation has shown that the best results for eye center detection can be achieved using SVM technique. The recognition rate increases statistically by 28% using face orientation normalization based on the eyes position.     

基于特征参数归一化的鲁棒语音识别方法综述

目前,自动语音识别系统往往会因为环境中复杂因素的影响,造成训练环境和测试环境存在不匹配现象,使得识别系统性能大幅度下降,极大地限制了语音识别技术的应用范围。近年来,很多鲁棒语音识别技术成功地被提出,这些技术的目标都是相同的,主要是提高系统的鲁棒性,进而提高识别率。其中,基于特征的归一化技术简单而有效,常常被作为鲁棒语音识别的首选方法,它主要是通过对特征向量的统计属性、累积密度函数或功率谱的归一化来补偿环境不匹配产生的影响。该文主要对目前主流的归一化方法进行介绍,其中包括倒谱矩归一化方法、直方图均衡化方法以及调频谱归一化方法等。     

虹膜图像预处理技术的研究

图像预处理是虹膜识别过程中至关重要的步骤。针对虹膜图像的特点和现有虹膜识别算法运算速度慢及鲁棒性差的问题,本文提出基于Canny算子和Hough变换的虹膜定位算法,然后采用Daugman的橡皮纸模型对虹膜区域进行归一化并做直方图均衡化图像增强处理。实验结果表明,本文的虹膜图像预处理算法可以有效解决图像预处理中计算时间长和通用性差的问题,能提高虹膜识别系统的稳定性和识别率。     

Radiometric normalization and image mosaic generation of ASTER thermal infrared data: An application to extensive sand sheets and dune fields

Data from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) have a significant advantage over previous datasets because of the combination of high spatial resolution (15-90 m) and enhanced multispectral capabilities, particularly in the thermal infrared (TIR) atmospheric window (8-12 μm) of the Earth where common silicate minerals are more easily identified. However, the 60 km swath width of ASTER can limit the effectiveness of accurately tracing large-scale features, such as eolian sediment transport pathways, over long distances. The primary goal of this paper is to describe a method for generating a seamless and radiometrically accurate ASTER TIR mosaic of atmospherically corrected radiance and from that, extract surface emissivity for arid lands, specifically, sand seas. The Gran Desierto in northern Sonora, Mexico was used as a test location for the radiometric normalization technique because of past remote sensing studies of the region, its compositional diversity, and its size. A linear approach was taken to transform adjacent image swaths into a direct linear relationship between image acquisition dates. Pseudo-invariant features (PIFs) were selected using a threshold of correlation between radiance values, and change-pixels were excluded from the linear regression used to determine correction factors. The degree of spectral correlation between overlapping pixels is directly related to the amount of surface change over time; therefore, the gain and offsets between scenes were based only on regions of high spectral correlation. The result was a series of radiometrically normalized radiance-at-surface images that were combined with a minimum of image edge seams present. These edges were subsequently blended to create the final mosaic. The advantages of this approach for TIR radiance (as opposed to emissivity) data include the ability to: (1) analyze data acquired on different dates (with potentially very different surface temperatures) as one seamless compositional dataset; (2) perform decorrelation stretches (DCS) on the entire dataset in order to identify and discriminate compositional units; and (3) separate brightness temperature from surface emissivity for quantitative compositional analysis of the surface, reducing seam-line error in the emissivity mosaic. The approach presented here is valid for any ASTER-related study of large geographic regions where numerous images spanning different temporal and atmospheric conditions are encountered.     

求解本原多项式的快速算法

文档图像处理技术是实现对网络上以"图片化"形式发送的垃圾邮件进行检测和过滤的有效手段.该文对彩色文档图像的版面进行分析,目的是分割出图像中的特定目标,便于分析并检测出文档图像中是否含有特别字符信息,从而使得网络垃圾邮件过滤系统可以根据这些信息判断是否过滤该邮件.实验结果表明,上述方法可以在不同颜色深度和不同几何结构的彩色文档图像中进行有效的检测,具有较好的实用性和应用价值.     

多姿态图像生成的行人重识别算法研究

不同行人的高度相似性以及相同行人外观姿态的差异性,使得不同摄像头下的行人重识别面临严峻的挑战。生成对抗网络可以合成新的图像,被认为是解决行人姿态变化的主要技术手段。提出一种基于多姿态图像生成的行人重识别算法,利用生成对抗网络生成不同姿态的行人图像,通过归一化消除姿态的影响,从而大幅度提升行人重识别的整体性能。该行人重识别算法包括多姿态行人图像生成、不同姿态的行人特征提取与融合、距离度量和重排序三部分内容。在Market-1501数据集和DukeMTMC-ReID数据集上的实验证实了所提出算法的有效性,通过与state-of-the-art行人重识别方法比较,展示了多姿态图像生成方法在行人重识别任务中的优越性,同时表明生成行人图像的特征与原始图像的特征是相互补充的。     

数据归一化方法综述

当今,人工智能已经广泛应用到各个领域中,并取得了显著的效果。数据归一化是人工智能应用落地中的一个重要环节,它有助于避免神经网络因数据量纲的复杂性对数据进行错误建模。在大数据场景下,相当一部分数据是以流的形式先后到达训练点,所以在流场景下数据归一化研究是当前亟待解决的关键问题。目前关于归一化研究的综述较多,大多仅仅针对于批数据的归一化研究,而缺乏对流数据的归一化方法的总结,不具参考性。在批数据归一化研究基础之上,系统化整理并详尽分析了流数据归一化的相关文献,凝练提出了基于流数据的归一化分类方法,并将数据归一化方法划分为批数据的归一化方法和流数据的归一化方法。同时,对这些方法的原理、优势和可以解决的主要问题进行了对比分析,在不同场景下对数据归一化的未来研究方向进行了展望。     

基于DeblurGAN的运动模糊图像盲复原算法研究

DeblurGAN方法利用条件生成对抗网络解决了端到端的图像去模糊问题,但存在图像边缘细节恢复不足以及鲁棒性不高的问题,针对此问题,提出一种基于DeblurGAN的运动模糊图像盲复原方法。在生成网络中,采用多尺度卷积核神经网络提取特征,并使用级联空洞卷积扩大神经元的感受野;采用自适配归一化方法代替原来生成器中使用的实例归一化方法。其次,引入了梯度图像L1损失,结合对抗损失和感知损失,将其作为图像去模糊的正则约束,使得生成图像的边缘特征更加清晰。实验结果表明,提出方法复原的图像峰值信噪比数值较DeblurGAN算法高出5.4%,结构相似性指标高出1%;在主观上清晰化效果较好,且消除了网格效应。     

Remote sensing of structural complexity indices for habitat and species distribution modeling

Spatial distribution models are increasingly used in ecological studies, but are limited by the poor accuracy of remote sensing (RS) for mapping microhabitat (< 0.1 ha) features. Mapping accuracy can be improved by combining advanced RS image-processing techniques with microhabitat data expressed as a structural complexity index (SCI). To test this idea, we used principal components analysis (PCA) and an additive SCI method developed for forest ecology (calculated by re-scaling and summing representative structural variables) to summarize 13 microhabitat-scale (0.04 ha) vegetation structure attributes describing the rare mountain bongo antelope's (Tragelaphus eurycerus isaaci) habitat in Kenya's Aberdare mountains. Microhabitat data were collected in 127 plots: 37 related to bongo habitat use, 90 from 1 km-spaced grid points representing overall habitat availability and bongo non-presence. We then assessed each SCI's effectiveness for discerning microhabitat variability and bongo habitat selection, using Wilcoxon Rank Sum tests for differences in mean SCI scores among plots divided into 4 vegetation classes, and the Area Under the Curve (AUC) of receiver operating characteristics from logistic regressions. We also examined the accuracy of predicted SCI scores resulting from regression models based on variables derived from a) ASTER imagery processed with spectral mixture and texture analysis, b) an SRTM DEM and c) rainfall data, using the 90 grid plots for model training and the bongo plots as an independent test dataset. Of the five SCIs derived, two performed best: the PCA-derived Canopy Structure Index (CSI) and an additive index summarizing 8 structural variables (AI₈). CSI and AI₈ showed significant differences between 5 of 6 vegetation class pairs, strong abilities to distinguish bongo-selected from available habitat (AUCs = 0.71 (CSI); 0.70 (AI₈)), and predicted scores 60-110% more accurate than reported by other studies using RS to quantify individual microhabitat structural attributes (CSI model R² = 0.51, RMSE = 0.19 (training) and 0.21 (test); AI₈ model R² = 0.46, RMSE = 0.17 (training) and 0.19 (test)). Repeating the Wilcoxon tests and logistic regressions with RS-predicted SCI values showed that AI₈ most effectively preserved the patterns found with the observed SCIs. These results demonstrate that SCIs effectively characterize microhabitat structure and selection, and boost microhabitat mapping accuracy when combined with enhanced RS image-processing techniques. This approach can improve distribution models and broaden their applicability, makes RS more relevant to applied ecology, and shows that processing field data to be more compatible with RS can improve RS-based habitat mapping accuracy.