基于图像深度预测的景深视频分类算法

Depth of field videos classification based on image depth prediction

景深视频因高清、美观广受大众喜爱。然而,要从海量视频中检出此类视频十分困难。已有较多研究基于景深图像成像原理,开展景深像素分割算法研究,但难以直接应用于实际视频分类场景。本文针对景深视频类型,设计了可预测视频类型的深度网络。根据景深成像原理,各语义物体之间相对相机的景深深度存在一定的逻辑关系。为此提出以图像深度为指导,利用深度预测模块预测图像的景深深度信息,将其合并后输入至分类网络进行训练检测,以降低景深视频误检率,提升网络模型的性能。此外,针对现实需求中该领域有标数据较少,而不同数据集分布会降低性能的问题,设计了迭代式景深视频数据集收集方法,以较低的劳动成本快速收集所需要的视频数据,具有一定的实际应用价值。本文算法在快手线上的景深视频数据集中识别准确率达85.7%。

Depth of field videos are usually beautiful and are very popular.However,it is a problem to classify such videos.There are many research works on the principle of the depth of field and segmentation algorithms,but they are often difficult to be applied to real video classification scenarios.This paper proposes a deep classification network to directly classify a video based on the observation that semantic objects with different depths of field usually have different definitions.According to the depth of field imaging principle,we propose to use the depth map as a guide to reduce the false detection rate and then improve the performance of the network.Furthermore,we design an iterative method to collect the depth of field videos quickly at low cost.Experimental results show that our method outperforms the previous methods,reaching 85.7%in the Kuaishou depth of field videos dataset.