【作者】 何聪； 金连文； 刘丽娟；

【Author】 HE CONG~(1)),JIN LIANWEN~(2)),LIU LIJUAN~(3)) 1)2)3)(School of Electronic and Information Engineering,South China University of Technology,Guangzhou 510640)

【机构】 华南理工大学电子与信息学院；

【摘要】 本文首次将生物启发特征(BIF)引入到笑脸识别中,尝试了一种模拟灵长类动物视觉识别机制的高层次的仿真生物视觉的新方法。针对生物启发模型(BIM)中的区域最大化操作对噪声的敏感性和信息的不完整性,提出了一种新颖的四邻域联合最大化操作;同时,为降低生物启发模型提取原型模板时的随机性所造成的信息冗余现象,引入Adaboost特征选择环节保留关键信息、去掉无用或负面信息。本文分别在标准数据库CKACFEID和现实环境数据库GENKI上进行比较实验,实验结果显示:生物启发特征(BIF)和本文改进的生物启发特征(IBIF)比目前流行的Gabor特征、PHOG特征等具有更好的识别性能和推广性能,而且,在现实环境数据库的笑脸识别中,IBIF仅用200维(远少于BIF的1000维)的特征就获得了最好的识别准确率。总的来说,IBIF更适用于极具挑战性的、受多种多样因素影响的现实生活中的笑脸识别,而且在特征存储量和运算量方面亦具有优势。更多还原

【Abstract】 We introduce biological inspired features(BIF),which are high-level features that simulate the primate visual recognition mechanism,into smile recognition for the first time.In order solve the problem of the sensitivity to noise and incompleteness of information in the max pooling operation in biological inspired model(BIM),we propose a novel pooling operation that combines the max point and its 4-neiborhoods.We also introduce Adaboost feature selection to retain the critical information and remove the useless information,so as to reduce the redundancy caused by the randomness in extracting prototype patches.We design comparison experiments both on the public standard database, CKACFEID,and the real-world database,GENKI.The results demonstrate that biological inspired features(BIF) and improved biological inspired features(IBIF) are superior to Gabor and other state-of-the art features in terms of effectiveness with less number of features and computational cost.In a word,IBIF is more suitable for smile recognition in the challenging realistic environment.更多还原