基于改进NeXtVLAD的视频分类

Video classification based on improved NeXtVLAD

为提高长视频分类精度并减少训练时占用显存,提出基于NeXtVLAD改进的长视频分类模型。将人脸识别领域的ghost聚类中心思想迁移到视频分类,通过加入ghost聚类中心降低无关采样帧的干扰,提高模型识别准确率,针对长视频分类提出多尺度的视频帧采样方法。采用预训练模型ResNet50提取采样帧的深度特征,在训练时冻结特征提取网络参数,减少训练时的计算量。在VideoNet数据集的前100个类别上进行实验,实验结果表明,该模型与现有相关模型相比取得了更好的分类效果。

To increase the accuracy of long video classification and reduce the occupation of video memory during training,an improved long video classification model based on NeXtVLAD was proposed.The idea of ghost clustering center in the field of face recognition was transferred to video classification.By adding ghost clustering center to reduce the interference of irrelevant sampling frames,the accuracy of model recognition was improved.A multi-scale video frame sampling method was proposed for long video classification.The pre-training model ResNet50 was used to extract the depth feature of the sampling frame,and the feature was freezed to extract the network parameters during the training to reduce the calculation amount during the training.Experimental results show that the proposed model has better classification effects than the existing correlation model.