【摘要】：通過結合高光譜數(shù)據(jù)與卷積神經(jīng)網(wǎng)絡(CNN)實現(xiàn)小麥不完善粒(黑胚粒、蟲蝕粒及破損粒)的快速準確鑒別。實驗采集小麥正常粒(484粒)、黑胚粒(100粒)、蟲蝕粒(100粒)及破損粒(100粒)在493~1 106 nm的116個波段的高光譜圖像,每間隔5個波段抽取1個圖像,分別建立24個波段的訓練集,應用CNN建立不完善粒小麥的識別模型。實驗結果顯示,利用該識別模型,黑胚、蟲蝕和破損粒的識別率分別保持在94%、95%和92%以上。在上述工作的基礎上,進一步通過修改學習率和迭代次數(shù)改進CNN模型。優(yōu)化后,黑胚、蟲蝕及破損粒在各波段下的平均識別率分別提高了0.624%、0.47%和0.776%。將24個波段高光譜圖像混合重新構建訓練集,并重新訓練CNN模型,黑胚、蟲蝕及破損粒的總識別率則分別提高了0.31%、0.13%和0.46%。綜上所述,基于高光譜數(shù)據(jù)和改進CNN模型可以有效提高小麥不完善粒的識別精度。
[Abstract]:Combined with hyperspectral data and convolution neural network (CNN), the rapid and accurate identification of wheat imperfect grains (black embryo, wormwood and damaged grains) was achieved. The hyperspectral images of wheat normal grain (484), black embryo (100), wormwood (100) and damaged grain (100) at 493 ~ 1 106 nm were collected. The training sets of 24 bands were established, and the identification model of imperfect grain wheat was established by CNN. The experimental results show that the recognition rates of black embryo, insect erosion and damaged particles are over 95% and 92%, respectively. On the basis of the above work, the CNN model is further improved by modifying the learning rate and iteration times. After optimization, the average recognition rates of black embryo, insect erosion and damaged particles in each band were increased by 0.624% and 0.776%, respectively. The training set was reconstructed by mixing 24 band hyperspectral images and the CNN model was retrained. The total recognition rates of black embryo, insect erosion and damaged particles were increased by 0.31% and 0.46%, respectively. In conclusion, based on hyperspectral data and improved CNN model, the identification accuracy of wheat imperfect grains can be improved effectively.
【作者單位】： 北京工商大學計算機與信息工程學院食品安全大數(shù)據(jù)技術北京市重點實驗室;
【基金】：國家自然科學基金(61473009) 北京市自然科學基金(4174086)資助項目
【分類號】：TP183;TP391.41

【相似文獻】

相關期刊論文 前1條

1 張玉榮;陳賽賽;周顯青;王偉宇;吳瓊;王海榮;;基于圖像處理和神經(jīng)網(wǎng)絡的小麥不完善粒識別方法研究[J];糧油食品科技;2014年03期

，

本文編號：2427331