基于Ghost-SE-Res2Net的多模型融合语音唤醒词检测方法

语音唤醒词检测(WWD)是语音交互中的关键技术,选择合适大小的检测窗对WWD性能的影响很大。提出一种新的多模型融合方法,通过融合小检测窗和大检测窗的检测结果来提高WWD性能。多模型融合方法包含两个分类模型,分别使用小检测窗和大检测窗,均基于轻量化的挤压与激励残差网络(SE-Res2Net)模块,即GhostSE-Res2Net,SE-Res2Net结构的多尺度机制可显著提升WWD的能力。在Ghost-SE-Res2Net中,首先使用Ghost卷积替换SE-Res2Net中的普通卷积以降低模型参数量,然后使用注意力池化层替换SE-Res2Net中的全局平均池化层进一步提升WWD能力。在实际检测时融合连续3个小检测窗模型的检测结果的最大值和1个大检测窗模型的检测结果,来判断唤醒词是否被触发。在训练时引入困难样本挖掘算法,选择性地学习较难检测的唤醒词信息以提高分类模型的检测性能。在包含2个唤醒词的Mobvoi数据集上评估系统性能,实验结果表明,在每小时0.5次错误唤醒的情况下,该系统在2个唤醒词上的错误拒绝率分别为0.46%和0.43%,实现了与先进基线相似的性能,并且系统参数量比基线少31%。

基于Ghost-YOLOv5s的SAR图像舰船目标检测

基于星载合成孔径雷达(synthetic aperture radar,SAR)图像的舰船目标检测中,为了平衡模型大小与检测精度,提出了一种基于Ghost卷积的SAR图像舰船目标检测方法Ghost-YOLOv5s。在YOLOv5s的颈部引入Ghost卷积,以减少模型参数和压缩模型体积;将高效的通道注意力机制(efficient channel attention,ECA)融入到颈部的C3模块里,以突出重要特征,从而保持较高的检测性能;使用SIoU损失函数替换原来的CIoU损失函数,以减少预测框和真实框之间的偏差,提高检测算法精度。实验结果表明,在SSDD遥感数据集上,改进模型与YOLOv5s相比,模型参数量减少了6.28%,模型体积减小了6.21%,检测精度达到了98.21%,实现了模型大小与检测精度的平衡。

基于Ghost模块的农资图像文本检测算法及其应用

针对农资图像中文本的检测速度慢并且缺乏移动端的应用等问题,基于农资图像数据集,提出了一种基于Ghost模块的农资图像文本检测算法,该算法对DB网络进行改进,使用MobileNetv2网络来提取基础特征,引入多尺度特征融合模块来获得多层之间的特征融合,并采用可微分二值化后处理算法预测文本,使其能够快速地检测农资图像中的文本。该算法在农资图像数据集上的准确率基本达到了主流算法的标准,检测速度达18.6 img/s,参数量为2.99 M,具备轻量级的特征,将此算法部署到移动端设备上并成功运行。

Joint Authentication Public Network Cryptographic Key Distribution Protocol Based on Single Exposure Compressive Ghost Imaging

In the existing ghost-imaging-based cryptographic key distribution(GCKD)protocols,the cryptographic keys need to be encoded by using many modulated patterns,which undoubtedly incurs long measurement time and huge memory consumption.Given this,based on snapshot compressive ghost imaging,a public network cryptographic key distribution protocol is proposed,where the cryptographic keys and joint authentication information are encrypted into several color block diagrams to guarantee security.It transforms the previous single-pixel sequential multiple measurements into multi-pixel single exposure measurements,significantly reducing sampling time and memory storage.Both simulation and experimental results demonstrate the feasibility of this protocol and its ability to detect illegal attacks.Therefore,it takes GCKD a big step closer to practical applications.

High-quality ghost imaging based on undersampled natural-order Hadamard source

Improving the speed of ghost imaging is one of the main ways to leverage its advantages in sensitivity and imperfect spectral regions for practical applications.Because of the proportional relationship between image resolution and measurement time,when the image pixels are large,the measurement time increases,making it difficult to achieve real-time imaging.Therefore,a high-quality ghost imaging method based on undersampled natural-order Hadamard is proposed.This method uses the characteristics of the Hadamard matrix under undersampling conditions where image information can be fully obtained but overlaps,as well as deep learning to extract aliasing information from the overlapping results to obtain the true original image information.We conducted numerical simulations and experimental tests on binary and grayscale objects under undersampling conditions to demonstrate the effectiveness and scalability of this method.This method can significantly reduce the number of measurements required to obtain high-quality image information and advance application promotion.

High-visibility ghost imaging with phase-controlled discrete classical light sources

We take phase modulation to create discrete phase-controlled sources and realize the super-bunching effect by a phasecorrelated method. From theoretical and numerical simulations, we find the space translation invariance of the bunching effect is a key point for the ghost imaging realization. Experimentally, we create the orderly phase-correlated discrete sources which can realize high-visibility second-order ghost imaging than the result with chaotic sources. Moreover, some factors affecting the visibility of ghost image are discussed in detail.

基于Ghost卷积的高级别浆液性卵巢癌复发预测方法

高级别浆液性卵巢癌是一种恶性肿瘤疾病,进行术前复发预测能帮助临床医生为患者提供个性化治疗方案,降低病人的死亡率。因该疾病的医学数据较少且难以获取,导致其深度学习模型难以得到充分的训练,复发预测准确率有待提高。针对此问题,本文设计了一种改进的低参数残差网络TGE-ResNet34,以ResNet34为主干网络,将传统卷积模块用Ghost卷积代替,完成病灶区特征的提取,降低模型的参数量,在2个Ghost卷积之间融入ECA(Efficient Channel Attention)注意力机制,抑制无用特征提取的干扰,最后通过5折交叉验证模型,避免数据随机划分的偶然性。实验结果表明,改进设计的TGE-ResNet34网络准确率为96.01%,相比原基线网络准确率提高4.52个百分点,参数量减少15.98 M。

Material Analysis of Traditional Folk Dwellings and Its Modern Inheritance:A Case Study of Traditional Folk Dwellings in Hunan Province

Traditional folk dwellings contain rich cultural connotations and plain architectural techniques.In architecture,material is the most fundamental thing,different materials can demonstrate different architectural forms,and reflect local characteristics and change of the time.Therefore,it is of great significance to explore the material selection of traditional Chinese folk dwellings.The paper took traditional folk dwellings in Hunan for example to analyze the regional materials and construction of these dwellings,discussed the application of traditional materials in modern architecture,used some cases to explore the innovative application of traditional materials,so as to figure out the new direction of applying traditional materials,and provide references for the construction of modern architecture.

Relation and Relationships Between Folk Literature and Classical Literature:Taking Autochthony as a Case Study

According to current discipline classification,folk literature is usually categorized as an individual research field,distinguished from classical literature or classical canon,which leads to the fact that in current research,the study of folk literature is often separated from the study of classical texts,or even opposed to it.This paper intends to review the relations and relationships between the two and argues that,at least,in ancient Greek world of culture,there is a considerable intersection between folk literature and classical literature.In many cases,a myth appeared both in the two literary fields,and thus these two fields were often intertwined and in conversation.Taking Athenian autochthony as a case study,this paper shows that this myth not only appeared as a folk story in daily occasions such as symposium,architecture,and vase painting,but also entered classic texts such as the tragic works of the three great tragedians and even Plato’s philosophical works,becoming one part of“classical literature”.By examining the application of the myth in folk literature and classical literature,it can be seen that in ancient Greece,folk and classical literature not only communicated and borrowed from each other,but also formed a dialogue between the two,thus constructing a lively and living literary world.

The Importance of Breathing Training in Folk Dance Teaching

The importance of breathing training in dance teaching is reflected in the two aspects of enhancing the quality of dance movements and sublimating the connotation of dance movements.For example,high-quality breathing can help performers complete the dance movements and improve the coordination of the movements;at the same time,the unique body rhythm formed by breathing can strengthen the visual effect of the performance and convey its spirit and soul to the audience.This requires folk dance teachers to carry out relevant training and teaching activities based on the categories and skills of dance breathing,such as changing students’ideological cognition,developing periodic breathing training courses,providing personalized guidance to students,and allowing students to adjust their learning and practice methods in the evaluation.

The Impact of Contextual Teaching Method on Sichuan Folk Song Education and Students’Musical Expressiveness

This study explores the application of the contextual teaching method in Sichuan folk song education and its impact on students’musical expressiveness.By incorporating contextual teaching methods in music classes,this research investigates the effectiveness of this approach in enhancing students’understanding of Sichuan folk songs and improving their musical expressiveness and emotional expression.A mixed-method research approach is employed,utilizing classroom observations,questionnaires,interviews,and statistical analysis to assess the practical outcomes of contextual teaching in folk song education.

基于相位变换和GhostNet-门控循环单元的自动调制识别方法

针对信号调制方式低信噪比条件下识别准确率不高的问题,提出了一种由相位变换、GhostNet、压缩与激励网络(Squeeze and Excitation Network,SENet)、门控循环单元(Gated Recurrent Unit,GRU)和深度神经网络组成的模型,用于自动调制识别接收信号。首先,采用基准数据集RML2016.10a和RML2016.10b同相正交数据作为模型输入;其次,构建识别模型,其中,相位变换用于降低相位偏移对调制识别的影响,GhostNet和GRU分别用于提取调制信号的空间特征和时间特征,SENet用于对特征图权重进行调整;而后,通过深度神经网络进行分类;最后,对所提模型进行了训练及测试。实验结果表明:与现有模型CGDNet、CLDNN、IC-AMCNet、MCLDNN和LSTM相比,所提出模型显著降低了参数量,有效提升了低信噪比条件下的识别准确率,平均识别准确率分别达到62.30%和64.45%。

基于Ghost模块的改进YOLOv5目标检测算法

现有以YOLOv5为代表的目标检测技术,存在骨干网络对特征提取不充分以及颈部层未高效融合浅层位置信息和深层高级语义信息等问题,这会导致检测精度较低,小目标误检、漏检。针对此问题,从兼顾实时性与检测精度出发,对YOLOv5进行改进,提出一种改进网络YOLOv5-CBGhost。首先在骨干网络中引入Ghost模块对模型进行轻量化处理,引入CA模块来更好地获得全局感受野,提高模型获取目标位置的准确度;然后借鉴双向加权特征金字塔网络,对原PAN结构进行改进,有效减少了特征冗余以及参数量,并通过跨层加权连接融合更多特征,提高了模型的目标检测精度;最后,增加多检测头以获取图片更丰富的高层语义信息,有效增加了检测精度。通过在PASCAL VOC2007+2012数据集上实验,YOLOv5-CBGhost的目标精度达到81.8%,相较于YOLOv5s,提高了3.0%,计算量减少42.5%,模型大小减少3.5%。

融合注意力机制与GhostUNet的路面裂缝检测方法

路面裂缝是道路最为常见的缺陷,随着深度学习技术的发展,利用深度学习的方法对路面图像中的裂缝信息提取的方法愈来愈多。针对现有深度学习路面裂缝检测方法提取裂缝特征不完整导致精度低以及实时性不足的问题,提出一种融合注意力机制与GhostUNet的路面裂缝检测方法。本方法由编码器和解码器组成,将U-Net中的常规卷积改进为Ghost卷积,减少模型参数量;在编码和解码部分,为了提高对裂缝特征的提取能力,引入ECA注意力机制和残差连接,ECA注意力模块可以过滤不相关的特征信息,利用残差连接可以避免网络退化现象。为评估本方法在裂缝检测方面的有效性,使用两个公开裂缝数据集,并进行消融实验和对比实验,实验结果F1_score、P和R分别比U-Net平均提高了14.48%、14.35%和14.45%;该模型相比U-Net参数量下降了14.2 MB。该模型与同类模型比较,分割的准确率更高,参数量更少。

Optimization method of Hadamard coding plate inγ‑ray computational ghost imaging

Owing to the constraints on the fabrication ofγ-ray coding plates with many pixels,few studies have been carried out onγ-ray computational ghost imaging.Thus,the development of coding plates with fewer pixels is essential to achieveγ-ray computational ghost imaging.Based on the regional similarity between Hadamard subcoding plates,this study presents an optimization method to reduce the number of pixels of Hadamard coding plates.First,a moving distance matrix was obtained to describe the regional similarity quantitatively.Second,based on the matrix,we used two ant colony optimization arrangement algorithms to maximize the reuse of pixels in the regional similarity area and obtain new compressed coding plates.With full sampling,these two algorithms improved the pixel utilization of the coding plate,and the compression ratio values were 54.2%and 58.9%,respectively.In addition,three undersampled sequences(the Harr,Russian dolls,and cake-cutting sequences)with different sampling rates were tested and discussed.With different sampling rates,our method reduced the number of pixels of all three sequences,especially for the Russian dolls and cake-cutting sequences.Therefore,our method can reduce the number of pixels,manufacturing cost,and difficulty of the coding plate,which is beneficial for the implementation and application ofγ-ray computational ghost imaging.

Ghost-Retina Net:Fast Shadow Detection Method for Photovoltaic Panels Based on Improved Retina Net

Based on the artificial intelligence algorithm of RetinaNet,we propose the Ghost-RetinaNet in this paper,a fast shadow detection method for photovoltaic panels,to solve the problems of extreme target density,large overlap,high cost and poor real-time performance in photovoltaic panel shadow detection.Firstly,the Ghost CSP module based on Cross Stage Partial(CSP)is adopted in feature extraction network to improve the accuracy and detection speed.Based on extracted features,recursive feature fusion structure ismentioned to enhance the feature information of all objects.We introduce the SiLU activation function and CIoU Loss to increase the learning and generalization ability of the network and improve the positioning accuracy of the bounding box regression,respectively.Finally,in order to achieve fast detection,the Ghost strategy is chosen to lighten the size of the algorithm.The results of the experiment show that the average detection accuracy(mAP)of the algorithm can reach up to 97.17%,the model size is only 8.75 MB and the detection speed is highly up to 50.8 Frame per second(FPS),which can meet the requirements of real-time detection speed and accuracy of photovoltaic panels in the practical environment.The realization of the algorithm also provides new research methods and ideas for fault detection in the photovoltaic power generation system.

基于YOLO v5的农田杂草识别轻量化方法研究

针对已有杂草识别模型对复杂农田环境下多种目标杂草的识别率低、模型内存占用量大、参数多、识别速度慢等问题,提出了基于YOLO v5的轻量化杂草识别方法。利用带色彩恢复的多尺度视网膜(Multi-scale retinex with color restoration, MSRCR)增强算法对部分图像数据进行预处理,提高边缘细节模糊的图像清晰度,降低图像中的阴影干扰。使用轻量级网络PP-LCNet重置了识别模型中的特征提取网络,减少模型参数量。采用Ghost卷积模块轻量化特征融合网络,进一步降低计算量。为了弥补轻量化造成的模型性能损耗,在特征融合网络末端添加基于标准化的注意力模块(Normalization-based attention module, NAM),增强模型对杂草和玉米幼苗的特征提取能力。此外,通过优化主干网络注意力机制的激活函数来提高模型的非线性拟合能力。在自建数据集上进行实验,实验结果显示,与当前主流目标检测算法YOLO v5s以及成熟的轻量化目标检测算法MobileNet v3-YOLO v5s、ShuffleNet v2-YOLO v5s比较,轻量化后杂草识别模型内存占用量为6.23 MB,分别缩小54.5%、12%和18%;平均精度均值(Mean average precision, mAP)为97.8%,分别提高1.3、5.1、4.4个百分点。单幅图像检测时间为118.1 ms,达到了轻量化要求。在保持较高模型识别精度的同时大幅降低了模型复杂度,可为采用资源有限的移动端设备进行农田杂草识别提供技术支持。

Optical encryption scheme based on spread spectrum ghost imaging

An optical encryption(OE) scheme based on the spread spectrum ghost imaging(SSGI), named as SSGI-OE, is proposed to obtain a high security with a smaller key. In the scheme, the randomly selected row number of a Hadamard matrix of order N is used as the secure key, and shared with the authorized user, Bob, through a private channel. Each corresponding row vector of the order-N Hadamard matrix is then used as the direct sequence code to modulate a speckle pattern for the ghost imaging system, and an image is encrypted with the help of the SSGI. The measurement results from the bucket detector, named as ciphertext, are then transmitted to Bob through a public channel. The illuminating speckle patterns are also shared with Bob by the public channel. With the correct secure key, Bob could reconstruct the image with the aid of the SSGI system, whereas the unauthorized user, Eve, could not obtain any useful information of the encrypted image. The numerical simulations and experimental results show that the proposed scheme is feasible with a higher security and a smaller key. For the 32 × 32 pixels image, the number of bits sent from Alice to Bob by using SSGIOE(M = 1024, N = 2048) scheme is only 0.0107 times over a computational ghost imaging optical encryption scheme.When the eavesdropping ratio(ER) is less than 40%, the eavesdropper cannot acquire any information of the encrypted image. The extreme circumstance for the proposed SSGI-OE scheme is also discussed, where the eavesdropper begins to extract the information when ER is up to 15%.

改进YOLOv7-tiny的手语识别算法研究

在与听障人士进行交流时,常常会面临交流不便的困难,文中提出一种手语识别的改进模型来解决这个困难。该模型基于YOLOv7-tiny网络模型,并对其进行了多项改进,旨在提高模型的精度和速度。首先,通过对CBAM注意力机制的通道域进行改进,解决了因降维而造成的通道信息缺失问题,并将改进后的CBAM加入到YOLOv7-tiny的Neck层中,从而使模型更加精准地定位和识别到关键的目标;其次,将传统的CIoU边界框损失函数替换为SIoU边界框损失函数,以加速边界框回归的同时提高定位准确度;此外,为了减少计算量并加快检测速度,还将颈部层中的普通卷积模块替换为Ghost卷积模块。经过实验测试,改进后网络模型的平均精度均值(mAP)、精准率和召回率分别提升了5.31%、6.53%、2.73%,有效地提高了手语识别网络的检测精确度。

Principle of subtraction ghost imaging in scattering medium

Scattering medium in light path will cause distortion of the light field,resulting in poor signal-to-noise ratio(SNR)of ghost imaging.The disturbance is usually eliminated by the method of pre-compensation.We deduce the intensity fluctuation correlation function of the ghost imaging with the disturbance of the scattering medium,which proves that the ghost image consists of two correlated results:the image of scattering medium and the target object.The effect of the scattering medium can be eliminated by subtracting the correlated result between the light field after the scattering medium and the reference light from ghost image,which verifies the theoretical results.Our research may provide a new idea of ghost imaging in harsh environment.