一种基于RBF网络生成人脸表情的算法

An efficient algorithm of facial expression generation based on RBF network

提出了只需确定少量标定点就能高精度实现人脸表情生成的算法。此方法并不直接利用标定点,而是通过自动生成的扭曲控制点来进行2D图像扭曲变形;为了保证精度,最大程度地减少人工输入标定点的数目。根据脸部肌肉的解剖特点和人脸的对称性定义了一个信息冗余很小的脸部特征点集(FCPS);通过径向基函数神经网络(RBFNetwork)来确定FCPS的移动与表情变化的映射关系;RBF网络输出的FCPS通过数据映射,自动生成一个矫正的扭曲控制点集;2D径向基函数图像扭曲算法(RBFWarpingAlgorithm)利用这些扭曲控制点扭曲图像,产生指定的表情,有效地降低了扭曲控制点误差带来的负面影响;为了避免病态数据的影响并减少计算量,训练网络时采用了递推正则最小二乘算法。实验显示,该算法具有良好的实用性。

The method to generate high-precision expression images by locating a small number of definite landmarks was proposed. The warping did not use landmarks as control points directly, but used a set of generated warping control points. Taking account of facial anatomy and symmetry, a novel set named Facial Characteristic Points Set(FCPS)was designed as landmarks with very low information redundancy. A radial basis function network (RBF Network) was employed to define the relationship between the movement of FCPS and the corresponding change of expressions. The FCPS, produced by RBF network, was processed by data-mapping component to generate a set of compensated warping control points. By using these generated warping control points, 2D RBF warping algorithm, which could effectively reduce the negative effects brought from the errors of generated warping control points, was employed to create assigned expression image. To prevent ill-posed data from affecting the computation and to reduce the computation complexity, recursive regularized least square (RRLS) algorithm was adopted in the neural network training.