晶圆缺陷检测与分类的卷积神经网络
晶圆缺陷检测与分类的卷积神经网络；针对晶圆检验时扫描电镜图像的缺陷检测和缺陷分类两问题，采用了“ ZFNet”的卷积神经网络来分类晶圆缺陷，并基于该分类器实现了一种“基于块的卷积神经网络”缺陷检测算法。为了提高准确率和加快速度，又改动“更快的区域卷积神经网络”实现了另一种检测算法。第卷第期邡鑫,史峥:晶圆缺陌检测与分类的卷积神经网络ZENet classifierDarker ImIn.ril” HumpBitel检测算法示意图在训练检测器时,数据集是检测器原始尺寸的图像,且包含标记好的缺陷区域和类型。我们结构通过·系列数据扩张操作,得到组数据,随机选取相比于检测算法主作为训练集,作为测试集。要从以下三方面进行了针对性的改进算法中需要优化的参数有滑动窗口尺寸滑()针对重复计算卷积的缺点,采用先动步幅、概率阙值、面积阙值,由于无法求出统一计算特征图,再按)进行映射各参数与检测结果的明确关系式,所以采用遍历法优化参截取的办法。如图,先通过卷积网络(数。因为检测到的缺陷尽量正确和尽量检测到所有缺陷是)对输入图像计算得到其特征图,因为在输入图像矛盾的,故以精确率和召回率的调和平均值作为优上的都能映射到特征图上,所以从输入图像上按化目标,也可根据实际需要调整两者权重满足不同侧重。割取图像进行卷积运算可以替代为直接从特征图上按测试结果映射后的范围割取,从而避免多次重复计算卷积。由于用训练好的检测模型对测试集检的大小形状不·,而全连接层的神经元连接数是固定的,测,计算模式下每张图大概耗时如果检测到的缺所以对割取得到的子特征图,通过层次采样到统陷与标准答案的且类型相同,则判为正确,否尺寸以连接到全连接层。则判为错误。得到结果如表,计算得:laut Image精确率Feature Map召回率ROI其屮正确缺陷的平均表检测器测试结果数量正确错误network有缺陷(正类)图映射示意图从检测结果来看该算法基本实现∫对图像上晶圆()针对滑动窗口尺寸单·的缺点,增缺陷的检测和分类,但是值较低,缺陷检测位置不加了滑动窗口的尺寸类型,并且增加由一个全卷积网络准确,检测耗时较长,分析其原囚如下)组成的()检测出错的数据中,缺陷较大的类型易判断错,)来预判断是否有缺陷。本文采用面积缺陷较小的容易被漏掉,说明只使用一种尺寸的滑动框很分别为,长宽比分别为、共难适应尺寸变化范围较大的缺陷种尺寸的滑动窗口,依次计算其中有缺陷的概率,再从中)滑动框步幅减小则算法耗时平方倍增加,而步幅筛选出一定数量最有可能有缺陷的区域,进行非极大值抑过长造成缺陷概率分布图分辨率较差,从而检测到缺陷位制(),最后得到一定数置准确度较差量的候选区域。()相邻滑动框都有大量重叠,所以每个区域都被多()针对缺陷检测位置准确度差的缺点,次重复送入计算卷积,导致算法耗吋较长。在全连接层后连接一个边界回归层在与上述检测算法相似的图像目标检测领域,近来出用来修正缺陷位置,该回归层与分类层并列。现的很好的克服了以上缺点并取得了很好的针对本文的缺陷检测问题,直接套用标准效果,所以下面介绍如何通过改动实现品圆并不能解决问题。因为判断晶圆的缺陷类型通常需缺陷的检测与分类。要结合缺陷区域周围的图形信息,而在预判断是否有C1994-2017ChinaAcademicJournalElcctronicPublishinghOusc.Allrightsrescrved.http://www.cnki.nct计算机工程年月日缺陷吋还进行了边界回归。虽然更加准确的给出缺陷的位()将原尺寸为的图像调整为置,但送入检测网络的特征儿乎不包含缺陷周围图肜信息,使得滑动窗口尺寸能够适应缺陷大小的变化范围,也可以导致缺陷分类不准。故木文对标准徹了一些根据实际情况来具体调整。改动:得到缺陷检测算法如图,卷积网络(()将改为只判断滑动窗口内是否有缺陷,而,)将输入图僚转换成多种特征图;根据不进行边界回归,也就是只计算所有滑动窗口有缺陷的概特征图从滑动窗口中选出最有可能存在缺陷的率,选取可能性最大的个,做非极大值抑制,再选出层根据特征图中抽取出对应特征组成特可能性最大的个进行检测。征向量;检测网络()根据特征向()将卷积层的尺寸加大为,加大感受野量判断缺陷类型,并进行边界回归;最后通过和概率),从而在判断滑动框內是否有缺陷吋能参阈值对候选缺陷进行过滤即可得到最终缺陷。考更多的周围信息。Detection NetworkonFolutionnl actorSoftmaxRuI Puling liver,e Prop卟 edMS+PrubilitessionInput Image 1024*1024Fully 10 dyercrectCcrvchrionalLaver size 747图检测算法示意图模型训练和平均值作为优化目标,并且使用相同的训练集和图中的检测算法也是基于架构实现,因为卷测试集积网络提取的特征类型对相似普遍有效,故其卷积网络的测试结果参数是直接迁移第章分类器的卷积层参数。但是用训练好的检测模型对测试集检测,和的参数则需要通过方法进行训练,标准计算模式下每张图大概耗时,采用相同判定标准,提供了分开和联合两种训练方式。为了节约得到检测结果如表(其中负类总数与表中总数不同是因时间,本文采用联合训练方式,并结合缺陷检测问题的实为同一张图屮可能检测到多个缺陷),计算得际情况调整超参数精确率在训练时,对每张输入图像,要计算的滑动窗口召回率数量庞大(种尺寸的滑动窗口,滑动步幅)。所以从中随机抽取个作为训练集,其中正例其中正确缺陷的平均负例,且正例占比不超过。分类器采用表检测器测试结果损失函数数量正确错误在训练时,设置提供个,从中随有缺陷(正类)机选取个作为训练集,其屮正例无缺陷(负类)负例,且正例占比不超过。另外设置学从结果来看该算法各方面都优于检测算习率分类器采用损失函数,而边界回法和值更高说明检测检测缺陷类型正确归采用函数。且位置准确,而且速度也大大提高(检测一张图像耗时从为了与检测算法对比,在最后通过遍历法缩小到)。如图为检测缺陷示例,共中标注了缺陷优化和概率阈值时,同样以精确率和召回率的调位置、类型和对应概率C1994-2017ChinaAcademicJournalElcctronicPublishinghOusc.Allrightsrescrved.http://www.cnki.nct邡鑫,史峥:晶圆缺陷检测与分类的卷积神经网络I I图检测结果示例图结束语而对图像上的缺陷检测和缺陷分类这两个问题,本文提出的改动后的检测算法能够精准、快速地从图像中检测出缺陷并同吋进行分类。得益于卷积神经网络良好的特征学习能力,该检测算法能够根据标记好缺陷位置和类型的数据自动学习特征,从而尽量避免人工千预,使算法具有较强的适应能力。参考文献徐姗姗刘应安徐昇基于卷积神经网络的木材缺陷识别山东大学学报工学版刘云杨建滨王传旭基于卷积神经网络的苹果缺陷检测算法电子测量技术江帆刘辉王彬等基于模型的图像识别计算机工程C1994-2017ChinaAcademicJournalElcctronicPublishinghOusc.Allrightsrescrved.http://www.cnki.nct
- 2021-05-06下载
- 积分：1
使用CC2591作为CC2530的功放
使用CC2591作为CC2530的功放, CC2591 PAThe absolute maximum ratings and operating conditions listed in the CC2530 datasheet [1]and the CC2591 datasheet [4] must be followed at all times. Stress exceeding one or more ofthese limiting values may cause permanent damage to any of the devicesNote that these characteristics are only valid when using the recommended register settingspresented in Section 4.6 and in Chapter 8, and the CC2530 - EM reference designOperating Frequency240524835MHzOperating Supply Voltage2036VOperating Temperature-40CTC=25C, VDD=3.0V, f=2440 MHz if nothing else is stated. All parameters are measuredon the CC2530-Cc2591EM reference design [11] with a 50 Q2 loadReceive CurrentWait for sync, -90 dBm input levelWait for sync, -50 dBm input level24mATXPOWER OXE5166mATXPOWER OXD5149mATXPOWER OXC5138mATXPOWER OXB5127mATransmit currentTXPOWER OXA5115ATXPOWER = 0X95100mATXPOWER = 0X8594ATXPOWE=0×75mATXPOWE=0×6579APower Down Current PM2UAISTRUMENTSPage 3 of 19SWRA308ATC=25C, Vdd=3.0V, f= 2440 MHz if nothing else is stated. All parameters are measuredon the CC2530-CC2591 EM reference design with a 50 Q2 loadReceive Sensitivity HGM 1 %PER, IEEE 802. 15.4[6] requires -85 dBm-988dBmReceive Sensitivity LGM1 PER, IEEE 802. 15.4 [6] requires -85 dBm-90.4dBmSaturationlEEE 802.15. 4 [6] requires-20 dBm10dBmWanted signal 3 db above the sensitivity levelIEEE 802.15.4 modulated interferer at ieee 802.15.4 channelsInterferer Rejection+5 MHz from wanted signal, IEEE 802. 15. 4 [6] requires 0 dBdB+10 MHz from wanted signal, IEEE 802. 15. 4 [6] requires 30 dB49dB+20 MHz from wanted signal wanted signal at- 82d BmdBdue to in the external lna and the offset in cc2530 the rssi readouts from cc2530CC2591 is different from rssi offset values for a standalone cc2530 design the offsetvalues are shown in table 4.4High Gain Mode79LoW Gain mode67Real rssi Register value-Rssl offsetISTRUMENTSPage 4 of 19SWRA308ATc=25C, Vdd=3.0V, f=2440 MHz if nothing else is stated All parameters are measuredon the CC2530-CC2591 EM reference design with a 50 Q2 load Radiated measurements aredone with the kit antennaRadiated Emissionwith TXPOWer Oxe5Conducted 2. RF (FCC restricted band)-462|dBmConducted 3. RF(FCC restricted band46.5 dBmComplies withFCC 15.247. SeeChapter 7 for moredetails about regulatoryRadiated 2.RF(FCC restricted band)42.2dBmrequirements andcomplianceIEEE 802.15.4[6]requires max.35%%Measured as defined by IEEE 802.15. 4 6TXPOWER OxE5. f= EEE 802.15. 4 channels13TXPOWER= OXD5. f= EEE 802.15.4 channelsTXPOWER= OXC5 f= EEE 802.15.4 channelsMax error∨ ectorTXPOWER OxB5 f= IEEE 802.15. 4 channelsMagnitude(EVM)TXPOWER OxA5. f= IEEE 802.15.4 channelsTXPOWER 0X95. f= IEEE 802. 15.4 channels643333%%%%%%%TXPOWER= 0x85. f= iEEE 802. 15.4 channelsTXPOWER =0x75 f= IEEE 802. 15.4 channels%TXPOWER= 065. f= iEEE 802. 15.4 channelsThe RF output power of the CC2530- CC2591 EM is controlled by the 7-bit value in theCC2530 TXPOWER register. Table 4.6 shows the typical output power and currentconsumption for the recommended power settings The results are given for Tc= 25 C, Vdd3.0V and f= 2440 MHz, and are measured on the cC2530-CC2591 EM reference designwith a 50 Q2 load. For recommendations for the remaining CC2530 registers, see Chapter 8 oruse the settings given by SmartRF StudioOXE520166OxD519149OxC18138OxB517127OxA5161150x95141000x8513940X75860x651079Note that the recommended power settings given in Table 4.6 are a subset of all the possibleTXPOWER register settings. However, using other settings than those recommended mightINSTRUMENTSPage 5 of 19SWRA308Aresult in suboptimal performance in areas like current consumption, EVM, and spuriousemissionTc=25C, Vdd=3.0V, f=2440 MHz if nothing else is stated All parameters are measuredon the CC2530-CC2591EM reference design with a 50 32 load2221-2V201918171611121314151617181920212223242526251510OxE5OxC5OxA50X850x65540-30-20-1001020304050607080ISTRUMENTSPage 6 of 19SWRA308A98Avg 3.6VAva 3vAvg 2V110111213141516171819202122232425261023.6V-1062V-110-40-30-20-100102030405060708070604020-Wanted signal at:-82 dBm10ISTRUMENTSPage 7 of 19SWRA308ACC2530-CC2591EM High Gain ModeC C2530-CC2591EM Low Gain Mode- CC2530EM40000-100110100908070-60-50-40-30-20-100The IEEE standard 802.15. 4 [8] requires the transmitted spectral power to be less than thelimits specified in table 4.7If-fc>3.5 MHz-20 dB-30 dBmThe results below are given for Tc=25 C, Vdd=3.0V and f= 2440 MHz, and are measuredon the CC2530-CC259 1EM reference design with a 50 Q loadIEEE absoluteChannel 182432.52435243752442524452447.5ISTRUMENTSPage 8 of 19SWRA308AOnly a few external components are required for the CC2530-CC2591 reference design. Atypical application circuit is shown below in Figure 5.1. Note that the application circuit figuredoes not show how the board layout should be done. The board layout will greatly influencethe RF performance of the CC2530-CC2591EM. TI provides a compact CC2530CC2591 EM reference design that it is highly recommended to follow. The layout, stack-upand schematic for the CC2591 need to be copied exactly to obtain good performance. Notethat the reference design also includes bill of materials with manufacturers and part numbersL102 L10=TI INF inductorVDD13cc2530LA 1RF PANTCC2591 RF NFNPA EN(P1 1)i工工I NA FNP:1HGM ENPO 7)T:1Proper power supply decoupling must be used for optimum performance. In Figure 5.1, onlythe decoupling components for the CC2591 are shown. This is because, in addition todecoupling, the parallel capacitors C11, C101, and C131 together with, L101, L102, TL11TL101 and TL131 also work as RF loads. These therefore ensure the optimal performancefrom the CC2591. C161 decouples the AvDD blAs power.The placement and size of the decoupling components, the power supply filtering and thePCB transmission lines are very important to achieve the best performance Details about theimportance of copying the CC2530-CC2591EM reference design exactly and potentialconsequences of changes are explained in chapter 6The RF input/output of CC2530 is high impedance and differential. The CC2591 includes abalun and a matching network in addition to the PA, LNa and RF switches which makes theinterface to the CC2530 seamless. Only a few components between the CC2530 andCC2591 necessary for RF matching For situation with extreme mismatch(VSWR 6: 1 till 12: 1out-of-band as shown in Figure 6.2) it is recommended to include all the components asshown in Figure 5.1ISTRUMENTSPage 9 of 19SWRA308ANote that the PCB transmission lines that connect the two devices also are part of the RFmatching. It is therefore important to copy the distance between the devices, the transmissionlines and the stack-up of the PCB according to the reference design to ensure optimumperformanceThe network between the CC2591 and the antenna(L111, C112, C111 C113 and L112matches the CC2591 to a 50 2 load and provides filtering to pass regulatory demands. C111also works as a dc-blockR151 is a bias resistor the bias resistor is used to set an accurate bias current for internaluse in the cc2591The TI reference design contains two antenna options. As default, the Sma connector isconnected to the output of CC2591 through a 0 Q2 resistor. This resistor can be soldered offand rotated 90 clockwise in order to connect to the PCB antenna, which is a planar invertedF antenna(PIFA). Note that all testing and characterization has been done using the SMAconnector. The PCB antenna has only been functionally tested by establishing a link betweentwo EMs. Please refer to the antenna selection guide [6] and the Inverted F antenna designnote [7 for further details on the antenna solutionsISTRUMENTSPage 10 of 19SWRA308A
- 2020-11-30下载
- 积分：1
