年龄跨度感知的多任务亲属关系验证

AGE SPAN-AWARE MULTI-TASK LEARNING FOR KINSHIP VERIFICATION

针对具有不同年龄跨度的亲属关系对象易在人脸上表现出不同程度的相似性的问题,提出一种年龄跨度感知的多任务学习方法。其核心是在多任务学习框架下,将具有不同年龄跨度的亲属关系验证分别看作一个学习任务,并在任务间共享参数,以达到利用更多判别信息的目的。另外,为了充分利用人脸不同局部区域所蕴含的遗传相似性,借助金字塔多层结构选择不同尺度的人脸区域。在两个亲属关系人脸图像数据库上的实验结果表明,年龄跨度感知的多任务学习方法具有较高的验证性能。

Aimed at the problem that visual entities with different age spans usually show different degrees of similarity,a new age span-aware multi-task learning method(AS-MTL)is proposed.The core of this method was to consider each type of age spans as one task and learn them at one time in the framework of multi-task learning,by sharing useful structures among tasks.The spatial structures of different types of age spans were simultaneously learned,which enabled our algorithm to utilize the discriminative information of samples with various similarities.Thus,the effect of age span on performance could be solved.To fully utilize the genetic similarity in different regions,face images were modeled by using a pyramid multi-level representation where local descriptors were extracted from several blocks at different resolution scales.Extensive experiments on two kinship datasets demonstrate the feasibility and effectiveness of the proposed algorithm.