基于残差和卷积神经网络的生猪图像分块压缩感知  

Block-based compressed sensing of pig images based on residual and convolutional networks

作　　者：汤先美 王春宇 闫顺丕 张立平[3] 周小波 焦俊[1] TANG Xianmei;WANG Chunyu;YAN Shunpi;ZHANG Liping;ZHOU Xiaobo;JIAO Jun(School of Information and Artificial Intelligence,Anhui Agricultural University,Hefei 230036,China;Anhui Xilejia Group Co.,Ltd.,Bozhou Anhui 233500,China;Anhui Academy of Agricultural Sciences,Hefei 230001,China)

机构地区：[1]安徽农业大学信息与人工智能学院,合肥230036 [2]安徽喜乐佳生物科技有限公司,安徽亳州233500 [3]安徽省农业科学院,合肥230001

出　　处：《东北农业大学学报》2023年第10期70-78,共9页Journal of Northeast Agricultural University

基　　金：安徽省重点研究与开发计划项目(2023n06020051)。

摘　　要：针对常规图像采样和压缩方法存在重建图像模糊问题,提出基于残差和卷积神经网络的生猪图像分块压缩感知模型RC-BCSNet(Block-based compressed sensing of pig images based on residual and convolutional networks)。RC-BCSNet是基于残差和卷积神经网络,采用采样、初始重建、深度重建三段式网络结构。首先通过卷积层自适应学习采样,再进行初始重建图像,最后通过基于残差的卷积神经网络进行图像整体深度重建。结果表明,RC-BCSNet与3种不同经典分块压缩感知算法相比,在7个采样比下平均峰值信噪比(PSNR)最大/最小增益分别为6.16和2.18 dB,平均特征相似度(FSIMc)最大/最小增益分别为0.083和0.037 dB,为信息中心数据处提供数据支持。Aiming at the problem of blurred reconstructed images in conventional image sampling and compression methods,a block-based compressed sensing of pig images based on residual and convolutional network(RC-BCSNet)was proposed.RC-BCSNet was a three-stage network structure based on residual and convolutional neural network,which adopted sampling,initial reconstruction and deep reconstruction.First,the convolutional layer adaptive learning sampling was used,then the initial image reconstruction was carried out,and finally the depth reconstruction of the whole image was carried out by the convolutional neural network based on residual.The results showed that RC-BCSNet's maximum/minimum gain of average peak signal-to-noise ratio(PSNR)under seven sampling ratios was 6.16 and 2.18 dB,respectively,compared with three different classical block compression sensing algorithms.The maximum/minimum gain of average feature similarity(FSIMc)was 0.083 and 0.037 dB,respectively,providing data assurance for the data section of the information center.

关 键 词：压缩感知 残差学习 卷积神经网络 生猪图像 

分 类 号：S828[农业科学—畜牧学] TP391[农业科学—畜牧兽医]