摘要
目的基于重建高光谱图像技术实现对酿酒高粱品种的实时快速识别。方法对分层回归网络(hierarchical regression network,HRNet)进行改进,得到残差注意力分层回归网络(residual attention-hierarchical regression network,RA-HRNet)。利用该网络进行重建高光谱图像,并在此基础上建立双向长短期记忆网络结合注意力机制(bi-directional long short-term memory-attention,BiLSTM-Attention)的酿酒高粱品种识别模型。以原始RGB数据作为重建高光谱图像网络的输入,将输出重建的光谱图像作为酿酒高粱品种识别模型的输入,以完成酿酒高粱品种识别。结果RA-HRNet相比HRNet,模型参数量[Params(M)]降低80.5%,模型计算量[floating point operations per second,FLOPS(G)]降低80.2%,峰值信噪比(peak signal-to-noise ratio,PSNR)提升16.9%,平均相对绝对误差值(mean relative absolute error,MRAE)降低31.8%,均方根误差值(root mean squared error,RMSE)降低19.1%;相比高光谱检测,重建高光谱检测效率提升95.8%;酿酒高粱品种识别模型的识别准确率最高可达95.1%。结论基于RA-HRNet重建高光谱图像网络结合BiLSTM-Attention模型可以实时快速识别酿酒高粱品种。
Objective To achieve real-time and rapid identification of sorghum varieties for liquor production based on reconstructing hyperspectral images technology.Methods The hierarchical regression network(HRNet)was improved to obtain the residual attention-hierarchical regression network(RA-HRNet).Using this network to reconstruct spectral images,and based on this,a bi-directional long short-term memory-attention(BiLSTM-Attention)for liquor sorghum variety recognition model was established.The original RGB data was used as input for the reconstruction spectral image network,and the output reconstructed spectral image was used as input for the sorghum variety identification model,completing the identification of sorghum varieties for liquor production.Results Compared to HRNet,the RA-HRNet model had reduced the number of parameters[Params(M)]by 80.5%,reduced the floating point operations per second[FLOPS(G)]by 80.2%,improved the peak signal-to-noise ratio(PSNR)by 16.9%,decreased the mean relative absolute error(MRAE)by 31.8%and reduced the root mean squared error(RMSE)by 19.1%.Compared to high spectral detection,the efficiency of reconstructed spectral detection had increased by 95.8%,the recognition accuracy of the sorghum variety identification model for liquor production could reach up to 95.1%at its highest.Conclusion The combination of RA-HRNet reconstruction spectral image network and BiLSTM-Attention model can quickly identify sorghum varieties for liquor production in real time.
作者
王俊
田建平
何林
胡新军
谢亮亮
杨海栗
陈满骄
WANG Jun;TIAN Jian-Ping;HE Lin;HU Xin-Jun;XIE Liang-Liang;YANG Hai-Li;CHEN Man-Jiao(School of Mechanical Engineering,Sichuan University of Science and Engineering,Yibin 644000,China;Key Laboratory of Brewing Biotechnology and Application of Sichuan Province,Yibin 644000,China)
出处
《食品安全质量检测学报》
CAS
2024年第1期65-73,共9页
Journal of Food Safety and Quality
基金
四川省科技厅项目(2023YFS0451)
四川轻化工大学研究生创新基金资助项目(Y2023078)。
关键词
酿酒高粱
重建高光谱
品种识别
算法研究
sorghum for liquor production
reconstructing hyperspectral
variety identification
algorithmic study
