属性的识别对物体的识别起到了比较重要的作用,例如人脸验证和场景识别.提高属性的识别率对后面基于属性特征的应用的正确率有很大的影响.近些年来,有些工作也开始关注于属性的学习,而很多的工作都是基于属性之间独立的假设,但在实际中很多的属性都是强相关的,例如没有胡子和女性,光头和头发的颜色;很多的工作忽略了类别之间的不平衡性,例如光头的样本比例可能只占样本的很小一部分.基于这2个观察,本文提出一种基于多任务的类别不平衡的人脸属性识别网络架构,该网络结构是由Dense net修改而来.该方法比以往的方法效果要好,一定程度上缓解了不平衡问题,且参数少,计算效率更高,在公开人脸属性数据集CelebA和LFWA上的实验验证了该方法的有效性.%The recognition of attributes plays an important role in object recognition, such as face verification, activity recognition in video. The improvement of attribute recognition will lead to better result for the application which uses these attributes. In recent years, there are some works focusing on the attribute learning. However, attributes have been considered to be independent in most works. In practice, we know that is not right and many attributes are strongly related, such as no beard and female, bald and hair color. In addition, most works neglect the unbalance of difference class samples, such as images of bald is a very small part of sam-ples. Based on the two observations, we propose a multi-task unbalanced facial attribute recognition framework using modified Dense net. To a certain degrees, the framework relieves the unbalanced problem. The proposed method outperforms other methods, has less parameters and runs faster. We demonstrate the effectiveness of our method on two challenging publicly available datasets.
展开▼
