目的 针对局部梯度编码算子(LGC)只能在固定大小邻域内提取图像纹理特征的不足,提出了一种非对称邻域LGC算子(AR-LGC)多尺度融合的表情特征提取方法。方法 首先,对归一化的表情图像进行Gauss滤波处理;然后,对图像进行分块,对每个子块图像中每一像素点,采用不同邻域大小的AR-LGC算子得到两个二进制序列,将两个序列作按位逻辑异或得到一个新的序列,对此序列进行编码,计算每个子块的直方图分布,级联各子块直方图构成人脸表情的特征;最后用SVM分类器进行表情分类识别。结果 该算法在JAFFE库和CK库上进行实验,分别取得了95.24%和96.83%的平均识别率,并与CBP(中心化二值模式)、LBP(局部二值模式)、LGC和AR-LBP(非对称局部二值模式)算法进行了比较,在JAFFE库的平均识别率分别比CBP、LBP、LGC、AR-LBP高5.6%、4.85%、3.71%、2.40%,在CK库的平均识别率分别比CBP、LBP、LGC、AR-LBP高3.66%、2.50%、2.17%、1.66%,实验结果表明,该算法可以较准确地进行人脸表情识别。结论 本文所提的表情特征提取方法通过融合不同梯度不同尺度子邻域间的强度关系,可以很好地表达图像的局部特征和全局特征,与典型的特征提取算法的对比实验也表明了本文算法的有效性,表明本文算法适用于静态人脸表情图像的识别。

Objective To overcome the deficiency of local gradient coding, which only extracts texture feature in neighborhoods of a fixed size, we propose a novel multi-scale local gradient coding fusion method based on asymmetric regions for feature extraction of facial expressions. Method A normalized face image is preprocessed by using a Gaussian filter to reduce the impact of noise. Then, the preprocessed expression image is divided into several blocks. For each pixel of each sub-block image, multiple and differently sized operators of local gradient coding based on asymmetric regions are used to obtain two binary sequences. These binary sequences are fused into a new binary sequence according to the logical XOR. The new binary sequence is then encoded, each sub-block histogram distribution is statistically analyzed, and all the sub-block histograms are cascaded into the texture features of a facial expression. Finally, the process of expression classification is completed with the SVM method. Result Experiments using the proposed method are performed using the JAFFE database and CK database. The average recognition rate for JAFFE is 95.24%, whereas that for CK is 96.83%. The proposed method is compared with LBP, CBP, LGC, and AR-LBP. Experimental results demonstrate that the proposed approach for the JAFFE database achieves recognition rates that are 5.6%, 4.85%, 3.71%, and 2.40% higher than those achieved with CBP, LBP, LGC, and AR-LBP, respectively. As for the CK database, the proposed approach achieves recognition rates that are 3.66%, 2.50%, 2.17%, and 1.66% higher than those achieved with CBP, LBP, LGC, and AR-LBP, respectively. Cross validation results show that the proposed method for facial expression recognition has excellent accuracy. Conclusion Through the fusion of the intensities between neighborhoods of different gradients and scales, the multi-scale local gradient coding fusion method based on asymmetric regions combined with block histograms can perform well in local and global feature description. Experiment results show that the proposed method is better than typical feature extraction algorithms and is suitable for static facial expression recognition.