基于非下采样shearlet梯度方向直方图和稀疏表示的脱机手写数字识别

《基于非下采样shearlet梯度方向直方图和稀疏表示的脱机手写数字识别》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、基于非下采样Shearlet梯度方向直方图和稀疏表示的脱机手写数字识别摘要：为了更有效的提高脱机手写体数字识别的性能和识别率，提山基于非下采样shearlet梯度方向直方图特征和稀疏表示的脱机手写字符识别方法。首先对字符图像进行非下采样Shearlet变换，得到低频子带图像和高频子带图像，然后将子图划分为若干矩形子块，分别统计子块区域的梯度方向直方图分布，将分块直方图串接起来形成非下采样梯度方向直方阁特征（HNSCOG）,最后用HNSCOH特征构建超完备字典，通过稀疏表示重构最小误差实现字符图像分类。在MN

2、IST和USPS数据集上测试，与DDH+SVM方法、SparseLS-SVM方法、sub-sampling+SVM方法和MTC+linear-SVM方法的识别率比较，实验结果表明，HNSCOG和稀疏表示的方法可以较大地提高脱机手写数字的识别率。I词：脱机手写数字识别，非下采样Shearlet变换，梯度方向直方图，稀疏表示Offlinehandwrittendigitcharacterrecognitionbasedonhistogramsofnonsubsampledshearletorientedgrad

3、ientsandsparserepresentationAbstract:Inordertoimproverecognitionrateofhandwrittendigitcharacterefficiently，anovelmethodofhandwrittennumeralscharacterrecognitionbasedonhistogramsofnonsubsampledshearletorientedgradientsfeatures(HNSCOG)andsparserepresentation

4、waspresentedinthispaper.Firstly，thenonsubsampledshearlettransform(NSST)wasutilizedtodecomposethecharacterimagesonvariousscalesandindifferentdirections，andthelowfrequencysub-bandandbandpasssub-bandcoefficientswereobtained.Then，sub-imagesofdigitalcharacterw

5、asdividedintogridsofblocksandcellstoextractHOGfeatures，andthehistogramofeachunitwascomputedandconcatenatedasHNSCOGfeaturesdescriptor.Finally，HNSCOGfeatureswerecombinedtoformanover-completedictionarywhichwasemployedbysparserepresentationtoclassifythehand

6、writtennumeralscharacter.Tocomparetheperformanceoftheproposedmethod，severalalternativealgorithmsforhandwrittennumberrecognitionhadbeenconsidered,forinstancedistancedistributionhistogram(DDH)plusSVMmethod，SparseLS-SVMmethod,sub-samplingplusSVMmethodandMTCp

7、luslinear-SVMmethod.Inordertoevaluatethesetechniques,acollectionofwellknownstandarddatabaseshadbeenused:MNISTdigitdatasetandUSPSdigitdataset.Theexperimentalresultsindicatethatthehandwrittennumeralcharacterrecognitionaccuracycanbeimprovedgreatly.Keywords:ha

8、ndwrittencharacterrecognition，nonsubsampledshearlettransform，histogramsoforientedgradients，sparserepresentation1引言脱机手写数字识别是光学字符识别领域一个具有挑战性的难题，有广阔的实际应用，如邮政信件分拣、表单数据处理等。联机手写数字识别已经取得较好的效果，然而脱机手写数字识别的精度离工程实践应川还存在