基于最优奇异值占比的融合特征人脸检测

Face detection with fusion features based on the proportion of optimal singular value

为了快速准确定位到人脸区域,解决人脸检测的速率和准确率有待进一步提高问题,提出了一种基于最优奇异值占比的融合特征人脸检测方法。首先,通过奇异值图像分解方法保留人脸图像有效信息,同时对图像实现降维重构;其次,在不同奇异值占比下,通过对比人脸图像的压缩率和准确率获取最优奇异值占比值;最后,基于最优奇异值占比,提出融合HOG,Haar特征人脸检测方法的集成分类方法。结果表明:在ORL人脸图像数据库上,获取的最优奇异值占比值为98%;人脸图像通过降维及重构处理后其空间复杂度降低了78.5%。在重构后的数据上,所提出的融合特征检测方法相对于HOG,LBP和Haar特征的人脸检测方法,检测准确率分别提高2%,17%和10%;相对于基于CNN的人脸检测方法其检测准确率降低0.5%,但检测速率提高99.2%。因此,作者提出的融合模型取得了较好的人脸检测效果。

In order to quickly and accurately locate the face area and solve the problem that the rate and precision of face detection are not satisfactory,a face detection method with fusion features based on optimal proportion of singular values is proposed.Firstly,the singular value image decomposition method is used to retain the effective information of the face image while achieving dimensionality reduction and reconstruction of the image.Secondly,in the case of different proportions of singular values an optimal one can be obtained by comparing the compression rate of the face image and the the accuracy of face detection.Finally,based on the optimal proportion of singular values,an integrated classification method fusing HOG and Haar feature face detection methods is designed.The results show that:on the ORL face image database,the obtained optimal proportion of singular values accounted for 98%;the space complexity of the face image was reduced by 78.5%due to dimensionality reduction and reconstruction procession.On the reconstructed data,the fusion feature detection method proposed in the article has improved detection accuracy by 2%,17%,and 10%compared to the face detection method with HOG,LBP,Haar features,respectively;the detection accuracy of the face detection method is reduced by 0.5%,but the detection rate is increased by 99.2%.Therefore,the fusion model put forward by the author has achieved better detection results.