摘要
在智慧城市、智慧巡检、智慧交通等场景中,摄像头等终端设备会产生大量的图像视频数据,并在云端由智能处理算法进行图形分析。然而传统的源端图像视频压缩传输,后端特征提取与分析识别的处理框架易造成视觉特征受损,影响分析识别精度。因此,源端提取图像特征,压缩后传输到后端的处理框架成为新的热点。文中提出了一种基于码率-准确率优化的图像特征压缩方法。首先,提取图像特征,分析划分特征图重要性的标准,将特征图分为重要性和非重要性特征两部分,并分别进行量化。在此基础上,建立码率-准确率的模型,在给定码率条件下,求解最优的准确率,确定相应的量化参数。以图像分类作为智能分析任务开展了实验。实验结果表明,所提出方法可以优化选择不同区域的量化参数,获得更好的编码性能。在低码率的条件下,相较JPEG算法准确率提高9.73%。
In smart city,smart inspection,smart traffic and other scenes,cameras and other terminal devices will generate a large amount of image video data that will be sent into the cloud and be analyzed by intelligent processing algorithms.However,these algorithms that usually deploy a processing framework including the traditional source-side image video compression transmission,as well as the back-end feature extraction,analysis and recognition can easily cause visual feature damages and affect the analysis and recognition accuracy.Therefore,the new mechanism that extracts the image features at the source side,compresses them and transmits them to the processing framework at the back end has become a hot topic.In this paper,we propose an image feature compression method based on rate-accuracy optimization.The image features are extracted.The criteria for dividing the importance of the feature map are analyzed.And the feature map is divided into two parts:importance and non-importance features,and quantified separately.On this basis,a rate-accuracy model is established,and the optimal accuracy is solved for a given bitrate condition to determine the corresponding quantization parameters.Finally,experiments are carried out with image classification as an intelligent analysis task.The results show that the proposed method can optimize the selection of quantization parameters for different regions and obtain better coding performance.The accuracy is improved by 9.73%compared to JPEG algorithm at a low bit rate.
作者
蒋伟
沈昊宇
JIANG Wei;SHEN Haoyu(School of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 200090,China)
出处
《南京邮电大学学报（自然科学版）》
北大核心
2024年第2期27-34,共8页
Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition
基金
国家自然科学基金青年基金(61401269)资助项目。
关键词
神经网络
特征压缩
分区域量化
码率-准确率优化
neural networks
feature compression
sub-regional quantification
rate-accuracy optimization
作者简介
蒋伟,女,博士,副教授,shiepjw@shiep.edu.cn。
