归类精度导引的在线图像集自适应压缩方法

Adaptive compression method for online image sets with classification accuracy preservation

针对多源视频流中的图像分类任务,提出了归类精度导引的在线图像集自适应压缩方法,首先对初始的在线图像集进行基于卷积神经网络(convolutional neural network,CNN)的模型训练,得到图像分类器;然后根据连续图像集之间的相似性,后一在线图像集的双参数参考前一在线图像集的双参数,通过引入自适应的参数判决机制有效地压缩连续的在线图像集。实验结果表明,所提方法能够保持足够大的平均压缩比,与现有的图像集压缩方法相比,可将平均归类精度提高3.3%。

Aiming at the image classification in multi-source video streams,this paper proposes an adaptive compression method for online image sets with classification accuracy preservation.The proposed method firstly performed the convolutional neural network(CNN)training on an initial online image set to obtain a classification model.Then,based on the similarity of continuous image sets,the double-parameter compression strategy of the latter online image set could refer to that of the previous online image set,and these continuous online image sets were effectively compressed by introducing an adaptive parameter decision mechanism.The experimental results show that the proposed method can maintain a large enough compression ratio,while the average classification accuracy can be improved by 3.3%as compared with the existing image set compression method.