融合改进韦伯特征的深度置信网络表情识别

Facial expression recognition based on binary Weber local descriptor and deep belief net

为改善深度置信网络运用于面部表情识别时,容易出现局部结构特征被忽视、鲁棒性差、运算量大等问题,提出融合双值韦伯局部描述子的深度置信网络算法。采用双值韦伯特征对图像进行初次特征提取,在空间分布优化传统韦伯特征的梯度方向算法,丰富局部细节纹理信息;在深度置信网络中进行二次特征提取,融合局部纹理信息的表征优势,借助深度学习在整体结构信息的提取优势,得到更易识别的高级抽象特征。实验结果表明,所提算法提高了面部表情识别率,减少了深度学习的计算量,对光照和噪音有更好的鲁棒性。

To solve the issues when deep learning was applied to facial expression recognition such as ignored local structure cha-racteristics,poor robustness of strong light and large computation.The method of facial expression recognition based on binary weber local descriptor(BWLD)and deep belief net(DBN)was proposed.First feature of image was extracted using BWLD.Compared with Weber local descriptor(WLD),orientation calculation method of local texture information in space distribution was optimized,details information was enriched.Second feature was extracted using DBN.Combined with the local texture feature advantage and whole structure information extracted advantage of deep learning,the higher level of more discriminating abstraction features was got.Experimental results show that,the algorithm of combination of DBN and BWLD results in higher recognition rate,and stronger robustness of illumination and noise.The amount of calculation of the deep learning is also reduced.