基于增强动态稠密轨迹特征的布料材质识别

为了克服基于视频的布料材质识别中由于忽略了动态因素的影响导致材质识别准确率低的难点,利用布料仿真视频库,提出一种基于增强动态稠密轨迹特征的布料材质识别方法以有效识别布料的材质属性信息.首先利用材质合成方法构造64种不同材质的布料仿真视频数据库;然后利用构造的布料动态视频库,通过迁移预训练的VGG网络增强布料动态视频的各帧特征信息,并消除部分非动态特征;其次利用动态稠密轨迹特征描述布料视频的动态特征,以充分捕捉不同布料材质视频的动态信息;最后通过Fisher向量编码生成布料动态信息的特征数据库,并训练SVM分类器建立布料材质视频动态信息到材质属性参数的映射.利用构建的布料仿真视频库作为数据集进行实验的结果表明,该方法对64种不同布料材质视频的材质种类识别准确率达到73.83%.

Material Recognition Using Enhanced Cloth Motion Dense Trajectory

In order to overcome the low recognition rate of different cloth materials due to neglecting the influence of the dynamic features of simulation video,this paper presents a novel cloth material recognition method using the enhanced cloth motion dense trajectory features.Firstly,a material synthesis method is presented to construct the simulation video database with 64 kinds of cloth materials.Then,the feature information of each cloth material video is enhanced and the non-dynamic features are eliminated by transferring the pre-trained VGG network.Secondly,in order to capture and represent the dynamic features of the cloth simulation videos with different materials,the novel cloth motion dense trajectory feature is calculated.Finally,the feature database of cloth dynamic information can be created by coding their fisher vectors,and the SVM classifier can also be trained to set up the mapping of dynamic information of cloth motion video to material attribute parameters.Experimental results show that the recognition rate for 64 kinds of different cloth materials is 73.83%by using our constructed cloth simulation video dataset.