利用高性能混合深度学习网络提升光谱分类性能研究

随着观测设备的不断完善,人们获得的光谱数量持续上升,如何进一步提高光谱自动分类的性能引起广泛关注.为此,以恒星光谱为研究对象,在近年来新出现的BERT和CNN等深度学习模型的基础上,试图融合了BERT模型和CNN模型在特征提取和智能分类方面的优势,提出高性能混合深度学习网络BERT-CNN,用以探讨该模型在提升光谱分类...     

Optimizing Few-Shot Learning Based on Variational Autoencoders

Despite the importance of few-shot learning, the lack of labeled training data in the real world makes it extremely challenging for existing machine learning methods because this limited dataset does not well represent the data variance. In this research, we suggest employing a generative approach using variational autoencoders (VAEs), which can be used specifically to optimize few-shot learning tasks by generating new samples with more intra-class variations on the Labeled Faces in the Wild (LFW) dataset. The purpose of our research is to increase the size of the training dataset using various methods to improve the accuracy and robustness of the few-shot face recognition. Specifically, we employ the VAE generator to increase the size of the training dataset, including the basic and the novel sets while utilizing transfer learning as the backend. Based on extensive experimental research, we analyze various data augmentation methods to observe how each method affects the accuracy of face recognition. The face generation method based on VAEs with perceptual loss can effectively improve the recognition accuracy rate to 96.47% using both the base and the novel sets.     

3E-Net: Entropy-Based Elastic Ensemble of Deep Convolutional Neural Networks for Grading of Invasive Breast Carcinoma Histopathological Microscopic Images

Automated grading systems using deep convolution neural networks (DCNNs) have proven their capability and potential to distinguish between different breast cancer grades using digitized histopathological images. In digital breast pathology, it is vital to measure how confident a DCNN is in grading using a machine-confidence metric, especially with the presence of major computer vision challenging problems such as the high visual variability of the images. Such a quantitative metric can be employed not only to improve the robustness of automated systems, but also to assist medical professionals in identifying complex cases. In this paper, we propose Entropy-based Elastic Ensemble of DCNN models (3E-Net) for grading invasive breast carcinoma microscopy images which provides an initial stage of explainability (using an uncertainty-aware mechanism adopting entropy). Our proposed model has been designed in a way to (1) exclude images that are less sensitive and highly uncertain to our ensemble model and (2) dynamically grade the non-excluded images using the certain models in the ensemble architecture. We evaluated two variations of 3E-Net on an invasive breast carcinoma dataset and we achieved grading accuracy of 96.15% and 99.50%.     

基于卷积神经网络的恒星光谱自动分类方法

恒星光谱自动分类是研究恒星光谱的基础内容,快速、准确自动识别、分类恒星光谱可提高搜寻特殊天体速度,对天文学研究有重大意义。目前我国大型巡天项目LAMOST每年发布数百万条光谱数据,对海量恒星光谱进行快速、准确自动识别与分类研究已成为天文学大数据分析与处理领域的研究热点之一。针对恒星光谱自动分类问题,提出一种基于卷积神经网络(CNN)的K和F型恒星光谱分类方法,并与支持向量机(SVM)、误差反向传播算法(BP)对比,采用交叉验证方法验证分类器性能。与传统方法相比CNN具有权值共享,减少模型学习参数;可直接对训练数据自动进行特征提取等优点。实验采用Tensorflow深度学习框架,Python3.5编程环境。K和F恒星光谱数据集采用国家天文台提供的LAMOST DR3数据。截取每条光谱波长范围为3 500~7 500 部分,对光谱均匀采样生成数据集样本,采用min-max归一化方法对数据集样本进行归一化处理。CNN结构包括：输入层,卷积层C1,池化层S1,卷积层C2,池化层S2,卷积层C3,池化层S3,全连接层,输出层。输入层为一批K和F型恒星光谱相同的3 700个波长点处流量值。C1层设有10个大小为1×3步长为1的卷积核。S1层采用最大池化方法,采样窗口大小为1×2,无重叠采样,生成10张特征图,与C1层特征图数量相同,大小为C1层特征图的二分之一。C2层设有20个大小为1×2步长为1的卷积核,输出20张特征图。S2层对C2层20张特征图下采样输出20张特征图。C3层设有30个大小为1×3步长为1的卷积核,输出30张特征图。S3层对C3层30张特征图下采样输出30张特征图。全连接层神经元个数设置为50,每个神经元都与S3层的所有神经元连接。输出层神经元个数设置为2,输出分类结果。卷积层激活函数采用ReLU函数,输出层激活函数采用softmax函数。对比算法SVM类型为C-SVC,核函数采用径向基函数,BP算法设有3个隐藏层,每个隐藏层设有20,40和20个神经元。数据集分为训练数据和测试数据,将训练数据的40%,60%,80%和100%作为5个训练集,测试数据作为测试集。分别将5个训练集放入模型中训练,共迭代8 000次,每次训练好的模型用测试集进行验证。对比实验采用100%的训练数据作为训练集,测试数据作为测试集。采用精确率、召回率、F-score、准确率四个评价指标评价模型性能,对实验结果进行详细分析。分析结果表明CNN算法可对K和F型恒星光谱快速自动分类和筛选,训练集数据量越大,模型泛化能力越强,分类准确率越高。对比实验结果表明采用CNN算法对K和F型恒星光谱自动分类较传统机器学习SVM和BP算法自动分类准确率更高。     

基于频率切片小波变换和支持向量机的癫痫脑电信号自动检测

实现癫痫脑电信号的自动检测对癫痫的临床诊断和治疗具有重要意义.本文提出先使用频率切片小波变换分离出5个不同频段的节律信号,再分别计算每个节律信号的近似熵和相邻节律的波动指数,最后使用遗传算法优化的支持向量机进行分类.实验结果表明,所提出的方法能够对正常、癫痫发作间期和癫痫发作期三种脑电信号进行准确分类,分类准确率为98.33%.     

水下高分辨率声图中小目标的深度网络分类方法

针对声成像数据缺少条件下的水下沉底小目标分类问题,提出一种深度网络分类算法。首先,采用高斯混合模型对声影区统计特性进行建模并提取声图阴影,在此基础上构建仿真数据集和真实数据集。将仿真数据集输入卷积神经网络进行训练,保留其特征提取部分,用于对真实数据集进行特征提取.重建网络分类部分并采用真实数据集的特征向量进行训练。结果表明,所提出的方法分类正确率可达88.24%,与6种对照方法相比平均分类正确率分别提升8.67%,20.47%,19.78%,11.59%,9.01%,11.58%。验证了所提出方法在小样本条件下具有较好对水下沉底小目标的分类能力。其学习曲线收敛到96.25%,仅比验证曲线高5.14%,说明在一定程度上缓解了过拟合问题。将改进的卷积神经网络应用于融合分类器,通过与逻辑回归分类器、支持向量机对目标进行分类并融合决策,正确率为93.33%,可进一步提高算法的正确率和稳定性.      

Wavelet basis selection and feature extraction for shift invariant ultrasound foreign body classification

This paper proposes a novel technique for automatic ultrasound non-destructive foreign body (FB) detection and classification. A signal registration process is introduced to eliminate shift variations commonly encountered in ultrasound signals. Information theory based methods are then developed for wavelet basis selection and feature extraction to facilitate robust FB classification. Probabilistic neural networks are used for FB classification. Experimental results confirm that the wavelet basis selected by the proposed method improves the FB classification accuracy. It is concluded that low order wavelet bases have better ability to distinguish classes with great similarities than their higher order counterparts, while the reverse is true for more divergent classes.     

结合光谱图像技术和SAM分类法的甘蓝中杂草识别研究

杂草自动识别技术是实现变量喷洒、精准施药的关键,更是制约其实现的瓶颈,因此,准确、快速、无损地实现杂草自动识别已成为精准农业的一个重要研究方向。利用高光谱成像系统采集甘蓝幼苗及小藜、稗草、牛筋草、马唐和狗尾草等五种杂草在1 000～2 500 nm波长区间的高光谱图像数据,在ENVI中经过MNF变换对数据降噪、去相关,并将波段维数从256维降到11维,通过提取感兴趣区域获得标准光谱,最后利用SAM分类法识别甘蓝与杂草,光谱角弧度阈值为0.1弧度时,分类效果良好。在HSI Analyzer中选择训练像元获得标准光谱后,利用SAM分类法识别甘蓝与杂草,并利用人工分类图与SAM分类图比较定量度量杂草的识别正确率,结果表明,当参数设置为5点平滑、0阶导数和7度光谱角度时,分类效果最佳,杂草识别率为80.0%,非杂草类识别率为97.3%,总体识别率为96.8%。应用光谱图像技术与SAM分类法相结合的方法进行杂草检测,充分利用了光谱和图像的融合信息,该方法应用空间的分类算法来建立光谱判别方法的训练集,在像素级别上考察光谱矢量之间的相似性,融合了光谱和图像两者的优势,同时兼顾了准确性和快速性,并且在整场范围内(行间和行内)改善杂草检测范围,为农业精确管理中需要植物精准信息的应用领域提供了相关的分析手段和方法。     

A Novel Method to Determine Basic Probability Assignment Based on Adaboost and Its Application in Classification

In the framework of evidence theory, one of the open and crucial issues is how to determine the basic probability assignment (BPA), which is directly related to whether the decision result is correct. This paper proposes a novel method for obtaining BPA based on Adaboost. The method uses training data to generate multiple strong classifiers for each attribute model, which is used to determine the BPA of the singleton proposition since the weights of classification provide necessary information for fundamental hypotheses. The BPA of the composite proposition is quantified by calculating the area ratio of the singleton proposition’s intersection region. The recursive formula of the area ratio of the intersection region is proposed, which is very useful for computer calculation. Finally, BPAs are combined by Dempster’s rule of combination. Using the proposed method to classify the Iris dataset, the experiment concludes that the total recognition rate is 96.53% and the classification accuracy is 90% when the training percentage is 10%. For the other datasets, the experiment results also show that the proposed method is reasonable and effective, and the proposed method performs well in the case of insufficient samples.     

基于可见光谱和卷积神经网络的贝类识别方法

目前卷积神经网络（CNN）在物体种类识别方面取得突破性进展。贝类作为农业经济的重要组成部分,种类繁多,特点复杂,大多贝类存在着相似度高,各类样本分布不均衡情况,以致CNN对贝类分类的准确率偏低。针对这一情况,提出了基于可见光谱和CNN的贝类识别方法,旨在提取更有效的贝类特征,从而提高贝类分类的准确率。首先,提出了一种包含输出熵度量和正交性度量的滤波器信息度量与特征选择方法,重新初始化修剪掉的滤波器并使其正交,捕获网络激活空间中的不同方向,使神经网络模型学习到更多有用的贝类特征信息,提升模型分类准确率;其次,提出了一种包含正则化项和焦点损失项的贝类分类目标函数,通过控制各类别样本对总损失的共享权重,来减少易分类样本的权重,以使模型注意力向预测不准的样本倾斜,均衡样本分布和样本分类难度,进一步提高贝类分类的准确率。贝类图像数据集由74类贝类组成,共11 803张图像。获取原始数据集后,对数据集图像进行水平翻转、垂直翻转、随机旋转、在[0, 30°]范围内旋转、在[0, 20%]范围内缩放和移动等数据增强操作,将图像数量从11 803张增加到119 964张。整个图像数据集按8∶1∶1的比例随机分为训练集95 947张图片、验证集11 996张图片和测试集12 021张图片。在建立贝类图像数据集的基础上进行了实验验证,达到了93.38%的分类准确率,将基准网络（Resnest）的准确率提高了1.18%,相较网络SN_Net和MutualNet,准确率分别提升了4.34%和0.85% ,并且训练时长为22 320 s,将基准网络（Resnest）的训练时长缩短了960 s,训练时长分别比SN_Net和MutualNet短3 180和2 460 s。实验结果证明了该方法的有效性。     

应用可见/近红外光谱进行黄酒品种的判别

为了实现对黄酒品种的快速判别,采用可见/近红外光谱对不同品种的黄酒获取光谱曲线,然后采用主成分分析方法对光谱数据进行聚类分析,并将其提取的主成分作为BP神经网络的输入值,建立了黄酒品种鉴别模型。该模型将前6个主成分作为神经网络的输入变量,加速了神经网络的学习速度,提高了模型的预测精度。随机选取每个品种的15个黄酒样本,共45个样本组成预测集,剩余的145个黄酒样本组成训练集建立训练模型,并用预测集样本对其进行验证。将品种鉴别的偏差标准定为±0.1,结果表明,只有1个未知样本超出偏差范围,该方法的品种鉴别正确率为97.78％,获得了满意的结果。说明文章提出的方法具有很好的分类和鉴别作用,为黄酒品种的快速鉴别提供了一种新方法。     

Bird Species Identification Using Spectrogram Based on Multi-Channel Fusion of DCNNs

Deep convolutional neural networks (DCNNs) have achieved breakthrough performance on bird species identification using a spectrogram of bird vocalization. Aiming at the imbalance of the bird vocalization dataset, a single feature identification model (SFIM) with residual blocks and modified, weighted, cross-entropy function was proposed. To further improve the identification accuracy, two multi-channel fusion methods were built with three SFIMs. One of these fused the outputs of the feature extraction parts of three SFIMs (feature fusion mode), the other fused the outputs of the classifiers of three SFIMs (result fusion mode). The SFIMs were trained with three different kinds of spectrograms, which were calculated through short-time Fourier transform, mel-frequency cepstrum transform and chirplet transform, respectively. To overcome the shortage of the huge number of trainable model parameters, transfer learning was used in the multi-channel models. Using our own vocalization dataset as a sample set, it is found that the result fusion mode model outperforms the other proposed models, the best mean average precision (MAP) reaches 0.914. Choosing three durations of spectrograms, 100 ms, 300 ms and 500 ms for comparison, the results reveal that the 300 ms duration is the best for our own dataset. The duration is suggested to be determined based on the duration distribution of bird syllables. As for the performance with the training dataset of BirdCLEF2019, the highest classification mean average precision (cmAP) reached 0.135, which means the proposed model has certain generalization ability.     

基于形态学成分分析的静态极光图像分类算法

为了解决海量极光图像手工分类效率低下的问题,提出一种静态极光图像自动分类系统,使用形态学成分分析将极光纹理从复杂背景中分离出来,从纹理中提取特征并利用支持向量机进行分类.实验结果表明:该算法分类正确率较之于传统方法均提高约10%,当分类器支持向量机+线性核函数时,分类速度最快,最适合于海量数据的处理.     

一种基于2D-CNN的激光超声表面缺陷检测方法

激光超声表面缺陷检测的过程中,缺陷的定量表征通常依赖于操作者的判断,易受到人为因素干扰,致使检测结果不稳定。针对这一问题,提出一种基于图像识别的二维卷积神经网络（2D-CNN）的缺陷自动分类检测方法。利用有限元方法模拟激光超声检测过程,并采集超声信号数据用于训练分类模型;使用连续小变换（CWT）处理超声信号得到小波时频图,以小波时频图作为输入训练卷积神经网络（CNN）分类模型,实现对表面缺陷深度的自动分类。验证结果表明:提出的检测方法能够对不同深度的缺陷准确分类,测试的平均准确率达到97.3%;构建的CNN分类模型能够自主学习输入图像的缺陷特征并完成分类,提高了检测结果稳定性,为激光超声缺陷检测的自动化分析处理提供了新的思路。     

Semi-supervised Discriminant Analysis and Sparse Representation-based self-training for Face Recognition

In this paper, we consider the problem of automatic face recognition with limited manually labeled training data. We propose a new semi-supervised self-training approach which is used to automatically augment the manually labeled training set with new unlabeled data. Semi-supervised Discriminant Analysis is used in each iteration of self-training for discriminative dimensionality reduction by making use of both labeled and unlabeled training data. Sparse representation is applied for classification. Experimental results on four independent databases show that our algorithm outperforms other face recognition methods under 3 different configurations, namely transductive, semi-supervised and single training image.     

Development of a noise sources classification system based on new method for feature selection

Feature selection of noise sources is important for noise sources detection and classification. In this paper, a new rough set based feature selection method has been given. Based on the method, a noise sources automatic classification system (NSACS) has been designed and validated. The key idea of the method is that most effective features can distinguish the most number of samples belonging to different classes of noise sources, if they are used for classification. This new approach has been applied into the system NSACS to select relevant features for artificial datasets and real-world datasets and the results have shown that this approach can correctly select all the relevant features of artificial datasets and at the same time it can drastically reduce the number of features. From the experiments, it can be found that to consider all the five datasets, the number of classification features after selection drops to 35% and the accurate classification rate increases about 14%. For the underwater noise sources dataset the number of features drops to 1/5 and the accurate classification rate increases about 6% after feature selection.     

DFTSA-Net: Deep Feature Transfer-Based Stacked Autoencoder Network for DME Diagnosis

Diabetic macular edema (DME) is the most common cause of irreversible vision loss in diabetes patients. Early diagnosis of DME is necessary for effective treatment of the disease. Visual detection of DME in retinal screening images by ophthalmologists is a time-consuming process. Recently, many computer-aided diagnosis systems have been developed to assist doctors by detecting DME automatically. In this paper, a new deep feature transfer-based stacked autoencoder neural network system is proposed for the automatic diagnosis of DME in fundus images. The proposed system integrates the power of pretrained convolutional neural networks as automatic feature extractors with the power of stacked autoencoders in feature selection and classification. Moreover, the system enables extracting a large set of features from a small input dataset using four standard pretrained deep networks: ResNet-50, SqueezeNet, Inception-v3, and GoogLeNet. The most informative features are then selected by a stacked autoencoder neural network. The stacked network is trained in a semi-supervised manner and is used for the classification of DME. It is found that the introduced system achieves a maximum classification accuracy of 96.8%, sensitivity of 97.5%, and specificity of 95.5%. The proposed system shows a superior performance over the original pretrained network classifiers and state-of-the-art findings.     

An improved automatic time-of-flight picker for medical ultrasound tomography

Objective and motivationTime-of-flight (TOF) tomography used by a clinical ultrasound tomography device can efficiently and reliably produce sound-speed images of the breast for cancer diagnosis. Accurate picking of TOFs of transmitted ultrasound signals is extremely important to ensure high-resolution and high-quality ultrasound sound-speed tomograms. Since manually picking is time-consuming for large datasets, we developed an improved automatic TOF picker based on the Akaike information criterion (AIC), as described in this paper.MethodsWe make use of an approach termed multi-model inference (model averaging), based on the calculated AIC values, to improve the accuracy of TOF picks. By using multi-model inference, our picking method incorporates all the information near the TOF of ultrasound signals. Median filtering and reciprocal pair comparison are also incorporated in our AIC picker to effectively remove outliers.ResultsWe validate our AIC picker using synthetic ultrasound waveforms, and demonstrate that our automatic TOF picker can accurately pick TOFs in the presence of random noise with absolute amplitudes up to 80% of the maximum absolute signal amplitude. We apply the new method to 1160 in vivo breast ultrasound waveforms, and compare the picked TOFs with manual picks and amplitude threshold picks. The mean value and standard deviation between our TOF picker and manual picking are 0.4 μs and 0.29 μs, while for amplitude threshold picker the values are 1.02 μs and 0.9 μs, respectively. Tomograms for in vivo breast data with high signal-to-noise ratio (SNR) (∼25 dB) and low SNR (∼18 dB) clearly demonstrate that our AIC picker is much less sensitive to the SNRs of the data, compared to the amplitude threshold picker.Discussion and conclusionsThe picking routine developed here is aimed at determining reliable quantitative values, necessary for adding diagnostic information to our clinical ultrasound tomography device - CURE. It has been successfully adopted into CURE, and allows us to generate such values reliably. We demonstrate that in vivo sound-speed tomograms with our TOF picks significantly improve the reconstruction accuracy and reduce image artifacts.