摘要
提出了一种基于自注意力机制增强的深度学习模型,用于无人机侦察图像的压缩与解压。与现有方法相比,提出的深度学习模型有两个显著特点:其一,模型由四部分组成(编码器、二值化器、量化器和解码器),并且可以通过端到端的优化提高模型的压缩和解压效率;其二,量化器是基于自注意力机制增强的多层前馈神经网络,它能充分利用图像的上下文信息对图像进行压缩。在公开数据集Kodak和Tecnick的实验结果表明,提出模型的压缩率-保真率曲线优于传统的图像压缩标准和现有的深度学习模型。对于常规大小的图像,在保持图像质量MS-SSIM为85%~95%的前提下,图像压缩比BPP能达到7%~15%,并且在普通CPU上其处理速度达0.48秒/张,能显著降低影像的数据大小且不牺牲处理速度。
A self-attention mechanism augmented deep learning model is proposed to compress and decompress the UAV reconnaissance image in this paper.Compared with the existing methods,the proposed deep learning model in this paper has two significant characteristics.Firstly,the model consists of four parts(encoder,binarizer,quantizer and decoder),and the compression and decompression efficiency of the model can be improved through end-to-end optimization.Secondly,the quantizer is a self-attention mechanism augmented multi-layer feedforward neural network,which can make full use of the context information to compress the image.Experimental results on public data sets such as Kodak and Tecnick show that the Bits Per Pixel-Peak Signalto Noise Ratial(BPP-PSNR)curve of the proposed model is better than that of the traditional image compression standards and existing deep learning models.For images with commonly used size,the compression ratio of the model,e.g.BPP,can reach 7%~15%while maintaining the MS-SSIM of 85%~95%,and the processing speed can reach 0.48 s/sheet on the ordinary CPU.The model proposed in this paper can significantly reduce the data size of the compressed image without sacrificing the processing speed.
作者
展亚南
施晓东
孙镱诚
丁阳
杨万扣
ZHAN Ya-nan;SHI Xiao-dong;SUN Yi-cheng;DING Yang;YANG Wan-kou(The 28th Research Institute of China Electronics Technology Group Corporation,Nanjing 210007,China;School of Automation,Southeast University,Nanjing 211189,China)
出处
《控制工程》
CSCD
北大核心
2022年第3期536-541,共6页
Control Engineering of China
基金
装备预先研究项目(301021302)。
关键词
图像压缩
深度学习
自注意力机制
端到端
多层前馈神经网络
Image compression
deep learning
self-attention mechanism
end-to-end
multi-layer feedforward neural network
