Automated identification of normoblast cell from human peripheral blood smear images

In this paper, we have presented a new computer‐aided technique for automatic detection of nucleated red blood cells (NRBCs) or normoblast cell from peripheral blood smear image. The proposed methodology initiates with the localization of the nucleated cells by adopting multilevel thresholding approach in smear images. A novel colour space transformation technique has been introduced to differentiate nucleated blood cells [white blood cells (WBCs) and NRBC] from red blood cells (RBCs) by enhancing the contrast between them. Subsequently, special fuzzy c‐means (SFCM) clustering algorithm is applied on enhanced image to segment out the nucleated cell. Finally, nucleated RBC and WBC are discriminated by the random forest tree classifier based on first‐order statistical‐based features. Experimentally, we observed that the proposed technique achieved 99.42% accuracy in automatic detection of NRBC from blood smear images. Further, the technique could be used to assist the clinicians to diagnose a different anaemic condition.     

一种基于模糊小波基函数神经网络的图像分类器

针对图像分类的特点，提出了一种基于模糊小波基函数神经网络的图像分类器。该分类器采用小波基函数作为模糊隶属函数，将模糊技术与神经网络相结合，利用神经网络实现模糊推理，并可对隶属函数的形状进行实时调整，从而使分类器具备了更强的学习和自适应能力。实验结果表明，这种基于模糊小波基函数神经网络的分类器经过训练后，可应用于图像的分类，其分类精度明显高于传统的最大似然分类法。     

Automated detection of tuberculosis in Ziehl-Neelsen-stained sputum smears using two one-class classifiers

Screening for tuberculosis in high-prevalence countries relies on sputum smear microscopy. We present a method for the automated identification of Mycobacterium tuberculosis in images of Ziehl-Neelsen-stained sputum smears obtained using a bright-field microscope. We use two stages of classification. The first comprises a one-class pixel classifier for object segmentation. Geometric transformation invariant features are extracted for implementation of the second stage, namely one-class object classification. Different classifiers are compared; the sensitivity of all tested classifiers is above 90% for the identification of a single bacillus object using all extracted features. The mixture of Gaussians classifier performed well in both stages of classification. This method may be used as a step in the automation of tuberculosis screening, in order to reduce technician involvement in the process.     

基于DRGAN和支持向量机的合成孔径雷达图像目标识别

为解决SAR图像目标识别中样本缺乏和方位角敏感问题,提出了一种基于DRGAN和SVM的SAR图像目标识别算法。首先,采用多尺度分形特征对SAR图像进行增强,经过分割得到目标二值图像,基于Hu二阶矩估计目标的方位角。然后对估计得到的目标方位角进行量化编码,结合原始图像作为输入,对设计的DRGAN模型参数进行训练与优化。由于DRGAN中的深度生成模型将目标姿态与外观表示进行解耦设计,故可利用该模型将SAR图像样本变换到同一方位角区间。基于变换后的训练样本分别提取归一化灰度特征,利用SVM训练分类器。采用MSTAR数据集在多个不同操作条件下对提出的算法进行测试,实验结果表明,在带变体的标准操作条件下,能够达到97.97%的分类精度,优于部分基于CNN模型的分类精度,在4种扩展操作条件下的分类精度分别为97.83%,91.77%,97.11%和97.04%,均优于传统方法的分类精度。在SAR图像目标方位角估计存在一定误差的情况下,训练得到的GAN模型作为SAR图像目标旋转估计器,能够使得在不进行复杂样本预处理的前提下,仍然取得较高的SAR图像目标识别精度。     

An improved initialization method of D-KSVD algorithm for bearing fault diagnosis

A novel bearing fault diagnosis method combining feature extraction based on wavelet packets quantifiers and pattern recognition method based on improved initialization method of Discriminative K-SVD (D-KSVD) algorithm is proposed. In D-KSVD algorithm, the representational power of dictionary and discriminative ability of classifier are seriously affected by their initialization values. Therefore, the improved initialization method of D-KSVD is presented and employed for bearing fault diagnosis. The improvement is that during the initialization of training stage, subdictionaries corresponding to each category are trained by K-SVD separately and then the initial dictionary is constructed by cascading the subdictionaries, which can completely represent the characteristics of all categories, and as for the initialization of linear classifier, naive Bayesian classifier is utilized. The experimental results show that under the same parameters the improved D-KSVD has better classification ability compared with traditional D-KSVD and some other classification methods.     

高光谱遥感图像空谱联合分类方法研究

在遥感影像研究领域里,高光谱数据分类是一个热点问题。近年来,在这个问题上涌现出很多研究方法,然而,大多数方法都是用浅层的方法提取原始数据的特征。将深度学习的方法引入高光谱图像分类中,提出一种新的基于深信度网络(DBN)的特征提取方法和图像分类架构用于高光谱数据分析。将谱域-空域特征提取和分类器相结合提高分类精度。使用高光谱数据进行实验,结果表明该分类器优于当前的一些先进的分类方法。此外,本文还揭示了深度学习系统在高光谱图像分类研究中具有的巨大潜力。     

结合噪声估计的NRLBP铆钉表面缺陷检测

针对铆钉表面缺陷纹理形态复杂多变,传统纹理特征提取方法难以获取准确纹理信息、铆钉缺陷识别率较低的问题,提出一种自适应阈值抗噪LBP的铆钉表面缺陷检测算法AT_NRLBP。首先,将铆钉图像均匀分块后提取铆钉子块;然后,基于PCA分解铆钉子块的协方差矩阵,估计子块图像的噪声水平。根据图像噪声强度计算NRLBP阈值,编码铆钉子块得到NRLBP纹理特征。最后,训练SVM单分类器分类铆钉子块,检测并标记出缺陷子块。实验结果表明,这里算法能有效地检测出铆钉表面缺陷,误检率明显降低;与其他纹理分类算法相比,这里算法在KTH-TIPS数据集上的分类准确率较高。     

基于双隐含层BP算法的激光主动成像识别系统

在传统激光主动成像系统的基础上,结合目标识别技术搭建了一个激光主动成像识别系统实验平台,用于研究激光主动成像后的目标识别。介绍了实验平台的工作原理,基于Hu矩特征的双隐含层BP神经网络算法以及实验处理流程和实验结果。特征量由7个不变Hu矩构成,通过240张原始目标样本库对由136个权值系数构成的双隐含层BP神经网络算法进行了训练。利用训练好的双隐含层BP算法对黑夜条件下远处的运动目标--43式冲锋模具枪进行了实验研究,成功获得了清晰的红外激光主动成像效果。实验显示对450 m处2 740帧和550 m处2 420帧激光主动成像图像的统计识别率达到了68.87%和72.11%,其中旋转变换下的统计识别率可达80.05%和84%,好于仿射变换的识别效果。     

基于遗传算法的支持向量机分类器模型参数优化

建立在统计学习理论和结构风险最小原则上的支持向量机在理论上保证了模型的最大泛化能力,因此与建立在经验风险最小原则上的神经网络模型相比,理论上更为完善。本文运用支持向量机建立模式识别分类器模型,研究影响模型分类能力的相关参数,在分析参数对分类器识别精度的影响基础上,提出用遗传算法建立支持向量机分类器模型的参数自适应优化算法。最后,用算例表明了本文算法的正确有效性。     

基于机器学习识别与标记分水岭分割的盲道图像定位

视觉导盲仪是一种旨在解决盲人出行困难的导盲设备,为了实现视觉导盲仪诱导盲人找到盲道并沿盲道行走,提出了一种基于机器学习识别与标记分水岭分割的盲道图像定位算法,通过离线训练与在线的识别、分割来定位盲道区域。首先对盲道图像进行视角变换的预处理,根据盲道的地面方程将变化的倾斜视角转换为固定的俯视视角,消除射影变换带来的失真;然后利用局部二进制模式描述子提取鸟瞰图的纹理特征,以自适应增强学习算法离线训练盲道识别分类器;进而利用分类器对鸟瞰图像进行在线识别,粗略确定盲道区域;最后将识别结果进行形态学处理后作为标记,利用标记分水岭算法得到精确分割的盲道区域并定位盲道中心线。在研制的视觉导盲仪上进行验证,结果表明盲道定位准确率达到了95.44%,速度平均每秒8帧,具有高准确率的同时达到实时性要求,为盲道的准确三维定位提供了必要条件。