【摘要】 该文针对新闻视频设计并实现了一个显著人脸检索系统。首先将新闻视频分割成镜头序列,利用训练好的CascadeAdaboost人脸检测器对每个镜头检测出一定数目的候选人脸,按照一些规则选取可信度高的作为样本,用于提取该镜头内的肤色模型。接着对肤色分割后的区域进行位置、大小分析和模板匹配,以淘汰非人脸区域,确定待跟踪的对象列表。为了做精确的跟踪和识别,系统对每个跟踪对象建立更细致的肤色模型。跟踪过程中每间隔一定帧数重新进行人脸检测,以减少误差积累和探测是否有新人脸出现。最后从每个人脸序列挑选最适合进行人脸识别的图像建立其特征脸空间,结合肤色信息和PCA算法判断其是否为要检索的目标人脸。更多还原

【Abstract】 A Dominant Face retrieval system for news videos is presented in this paper.Firstly,the news video is segmented into shot series,then a trained face detector is applied to each shot to get some candidate face regions,then subjected to rule -based selection,for the purpose to get highly creditable face regions as samples for skin color modeling of the current shot.After skin color segmentation,the candidate regions are checked against the reasonable position,size,and face templates.Passed regions are viewed as human faces and added into the tracking list,followed by the bidirectional face tracking.Correction will be periodically conducted after tracking every certain number of frames,to reduce the accumulated error and detect new faces timely.For higher accuracy in detection and later recognition,this paper constructs a specific skin color model,rather than the general shot color model,for each object in tracking list.When tracking finished,the system will try to find the best face image of each face series for recognition.At last,the PCA algorithm combined with individual skin color model comparison is used to determine which one is the target face.更多还原