基于生物特征质量的多模态稀疏表示自适应融合识别方法研究

Ideally, biometric performance can be promoted by combining multiple biometric traits, which however may not work due to the occurrence of data degeneration in applications. It remains open and challenging to develop effective biometric quality assessment and subsequent quality-based dynamic fusion of multimodal biometirc traits. Aiming to develop such adaptive multimodal recognition, this project will pay effort on the four aspects below. First of all, we propose to measure biometric sample quality in a 2D continuous space with the aid of sparse representation, based on the fact that whose reconstruction accuracy and coding distribution are strongly correlated with classification accuracy and hence are able to reflect the utility of biometric sample in recognition. Then, we attempt to design a fusion weight function for adaptively integrating sparsity-based match scores of each modality by taking into account unimodal samples’ qualities and the overall multimodal sample quality. This dynamic match score level fusion will strengthen the robustness of multimodal identification. Besides, a new multimodal joint sparse representation model is developed for better digging discriminative information at feature level when the biometric samples are of relatively good quality. As for a class in this model, while guaranteeing the similarity among coding vectors of modalities, we try to decrease the coupling degree among the corresponding coefficients of these vectors. Moreover, clustering analysis and dictionary optimization techniques are exploited to design dictionary that with a small scale so as to guarantee efficiency and accuracy of sparse coding for verification. An adaptive multimodal verification method can be finally achieved by taking advantage of multimodal joint sparse representation and biometric quality. Hopefully, this project will provide theoretical and technical and experimental support to adaptive multimodal biometric recognition.

融合多种类型生物特征可提高识别准确性，但实用中面临诸多干扰因素影响，如何有效评价样本质量并动态融合多模态生物特征亟待深入研究。为此拟从四方面着手：稀疏表示的样本重构和编码分布特性与分类性能强相关，能反映样本的识别可用性，拟构建基于稀疏表示的二维连续质量空间来描述生物特征质量；结合样本个体和多模态样本总体质量设计连续融合权重选择函数，进而动态融合多模态稀疏匹配分数，以提高身份辨识抗干扰能力；同时，对于样本质量较好情形，研究如何在多模态联合稀疏表示中，保障同一样本类的各模态编码向量总体一致的同时降低向量间各对应分量的耦合度，以便更合理和充分挖掘特征层丰富的判别信息；在身份验证方面，采用聚类分析和字典优化方法设计有限规模字典来保障稀疏表示身份验证的计算效率和可靠性，并设计稀疏匹配指标，研究结合生物特征质量的多模态自适应融合身份验证。本项目的开展将为多模态自适应融合识别提供技术支撑和数据依据。

多模态识别系统利用多种生物特征进行身份识别，虽可一定程度上提高识别精度和鲁棒性，但是任一模态生物特征样本质量下降都会影响多模态识别结果，若不能根据样本的质量趋利避害将会导致多模态识别效果不及单模态识的窘境。为此，本课题围绕着多模态自适应融合，从生物特征质量评价、自适应融合、多模态联合稀疏表示、多模态系统抗哄骗等方面开展研究，取得了一些列成果。（1）通过对多种类型图像噪声污染情况下的人脸、人耳和掌纹单模态识别实验，发现稀疏协作表示精度和稀疏编码的集中程度对不同类型噪声的敏感度不同，具有互补性，并且二者都与样本的识别可用性强相关，因此我们设计了分别基于二者的两种生物特征质量评价指标来描述样本的质量。（2）针对稀疏表示分类框架下匹配层多模态融合抗噪性好，但多模态联合稀疏表示能够更充分利用特征信息，提出基于生物特征质量的既能动态选择融合层次策略又能自适应融合多模态特征的身份辨识方法。（3）针对不同类型生物特征相关性低的特点，提出一种新的多模态联合稀疏表示模型，该模型引入了类别级编码一致和组稀疏正则项，使得各模态编码总体一致情况下更具灵活性。（4）研究基于稀疏表示的“一对多”竞争匹配身份验证的性能特点，设计基于编码稀疏性和重构精度互补性特点的竞争匹配度指标计算方法，并运用目标类和竞争类的协作表示精度来描述样本与目标类的相关度，最后结合各模态匹配度和相关度训练分类器，得到一种高抗哄骗多模态身份验证方法。大量的脸与掌纹、脸与耳的双模态验证实验结果表明，该方法能够有效提高系统抗哄骗能力的同时不降低普通验证情况下的精度，实现了二者良好的平衡。