深度学习轻量化侦察图像压缩网络  

Deep learning Based Lightweight Reconnaissance Image Compression Network

作　　者：谌宇 谌德荣[1] 崇魁奇 王泽鹏 张凯 CHEN Yu;CHEN Derong;CHONG Kuiqi;WANG Zepeng;ZHANG Kai(Beijing Institute of Technology,Beijing 100081,China;Beijing HYTQ Technology Ltd.Co.,Beijing 100043,China)

机构地区：[1]北京理工大学,北京100081 [2]北京航宇天穹科技有限公司,北京100043

出　　处：《探测与控制学报》2024年第1期78-84,共7页Journal of Detection & Control

摘　　要：为了满足小型侦察平台对低复杂度图像编码算法的应用需求,提出基于深度学习轻量化侦察图像压缩网络。轻量化侦察图像压缩网络编码端利用三个卷积模块直接将图像映射为服从均匀分布的二进制码流,得到压缩数据;在卷积模块中采用深度可分离卷积、分组卷积+通道重排等方式降低了编码端参数量和计算量。轻量化侦察图像压缩网络解码端采用转置卷积和残差连接等方式提高特征提取能力,进而提高解码图像质量。对分辨率为128×128实际采集图像的测试结果表明,与JPGE2000算法相比,基于深度学习轻量化侦察图像压缩网络PSNR提高了3.85 dB,编码时间降低了91%,实现了图像的轻量化编码压缩。In order to meet the application requirement of low complexity image coding algorithm for small reconnaissance platforms,a lightweight reconnaissance image compression network based on deep learning was proposed.At the coding end of the lightweight reconnaissance image compression network,three convolution modules were used to map the image directly to the binary code stream conforming to uniform distribution,and the compressed data was obtained.In the convolutional module,depth-separable convolution,group convolution plus channel shuffle were adopted to reduce the number of coding end parameters and the amount of computation.The decoder of lightweight reconnaissance image compression network applied to transposition convolution and residual connection to improve the ability of feature extraction and the quality of decoded image.Test results of 128×128 images showed that compared with JPGE2000,the PSNR of lightweight reconnaissance image compression network based on deep learning was increased by 3.85 dB,the coding time was reduced by 91%,and the lightweight coding compression of image was realized.

关 键 词：侦察图像压缩 深度可分离卷积 分组卷积 通道重排 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]