English

中国农机化学报

中国农机化学报 ›› 2023, Vol. 44 ›› Issue (10): 87-92.DOI: 10.13733/j.jcam.issn.2095-5553.2023.10.013

• 农产品加工工程 • 上一篇    下一篇

基于GoogLeNet的玉米籽粒破损及霉变在线辨识方法

林杰1, 2,王发赢1, 2,姚艳春1, 2, 3,崔春晓1, 2,盛振哲1, 2,曲殿伟3   

  1. 1. 山东理工大学农业工程与食品科学学院,山东淄博,255000;
    2. 山东省旱作农业机械及信息化重点实验室,山东淄博,255000;
    3. 山东五征集团有限公司,山东日照,276825
  • 出版日期:2023-10-15 发布日期:2023-11-09
  • 基金资助:
    十四五国家重点研发计划项目(2021YFD2000502);中国博士后基金资助项目(2022M711982)

Online identification method of corn kernel damage and mildew based on GoogLeNet

Lin Jie1,  2, Wang Faying1,  2, Yao Yanchun1,  2, 3, Cui Chunxiao1,  2, Sheng Zhenzhe1,  2, Qu Dianwei3   

  • Online:2023-10-15 Published:2023-11-09

摘要: 针对采集的玉米籽粒图像存在冗余信息多、目标区域占比小、目标位置具有随机性等问题,提出一种基于颜色空间(HSV)阈值分割的玉米籽粒图像切片处理方法,提高玉米籽粒破损及霉变的辨识准确率。首先,确定分割阈值,提取出玉米籽粒轮廓,根据轮廓的最小外接矩形框顶点坐标,确定裁剪坐标并裁剪图像;其次,获得未经图像切片处理与切片处理的GoogLeNet模型权重,在GoogLeNet第一、二、三层卷积及inception5b模块后使用GradCAM可视化方法,对不同卷积层所提取的特征进行可视化解释;最后,结合验证集准确率评价两组模型对完好、破损以及霉变玉米籽粒的特征提取能力。结果表明,本文提出的图像切片处理方法可将验证集准确率提高到93.74%,相较于未经图像切片处理的数据集,在GoogLeNet上验证集准确率提升7.99%;借助GradCAM可视化方法显示模型关注区域,可视化解释网络提取的特征,验证该方法的有效性,为玉米籽粒图像预处理及辨识分类提供新思路。

关键词: 玉米籽粒, 阈值分割, 破损, 霉变, 卷积

Abstract: Aiming at the problems such as excessive redundant information, small proportion of target area and randomness of target location in the collected corn grain images, a new corn grain image slicing method based on color space (HSV) threshold segmentation was proposed in this paper, which improved the recognition accuracy of damaged and mildewed corn   grain. Firstly, the segmentation threshold was determined, the corn grain contour was extracted, and the cropping  coordinates were determined and the image was clipped according to the vertex coordinates of the minimum external rectangle frame of the contour. Secondly, the weight of GoogLeNet model without image slicing processing and with image slicing processing were obtained. After the first, second and third convolution layer and inception5b module of GoogLeNet, the Grad-CAM visualization method was used to visualize the features extracted from different convolutional layers. Finally, based on the accuracy of verification set, the ability of the two models to extract intact, damaged and mildewed corn grain features was evaluated. The results showed that the method proposed in this paper could improve the verification set accuracy to 93.74%, which was 7.99% higher than that of the data set without image slicing processing on GoogLeNet. The models concerned areas were displayed by the Grad-CAM visualization method, and the features extracted from the network were interpreted visually, which verified the effectiveness of this method and provided a new idea for the pretreatment and identification of corn kernel image.

Key words: corn kernel, threshold segmentation, damaged, mildewed, convolution

中图分类号: