基于孪生非负矩阵分解的车脸重识别算法

Vehicle face re-identification algorithm based on siamese nonnegative matrix factorization

受光照强度变化影响,同一车辆在不同时段采集的车脸图像可能会存在差异,如车身颜色、车灯状态等,为了使识别方法对多种光照条件具有普适性,提出了一种孪生非负矩阵分解模型。首先,将每一对训练样本车脸图像的初始特征分配在两个非负矩阵分解模型中;然后,融合分解后的误差损失,类内损失,类间损失,设计了一种孪生非负矩阵分解模型,其中,两个非负矩阵分解模型共享同一特征基;最后,基于梯度下降法对模型进行求解,获得共享特征基,并基于余弦距离实现了车脸图像的匹配。实验结果表明,对于存在一定光照差异条件下采集的两幅车脸图像,提出的算法仍能获得较为准确的重识别结果,错误接受率与错误拒绝率均可降低至6%以下。

The light intensity variation may bring some differences among vehicle face images which are captured at different times such as vehicle color difference,headlight status difference,etc.To make the recognition method universal to multiple lighting conditions,a novel siamese nonnegative matrix factorization(NMF)model is formulated.First,the original features of each pair of vehicle face training images are split and taking as the input of two NMF models.Then,a siamese NMF model is established by fusing the error loss,the intra-class loss and the inter-class loss.The same feature basis vectors are shared by these two NMF models.Finally,the model is solved by using the gradient descent algorithm.Thus,the shared feature basis vectors can be acquired,and the re-identification of vehicle face images can be achieved based on the cosine distance.Experimental results show that the proposed algorithm can achieve accurate re-identification results even when two vehicle face images are captured under different lighting conditions.Both the false accept rate and the false reject rate can be reduced to be below 6%.