基于机器视觉和MSD微结构描述算法的霉变烟在线检测研究

Study on on-line detection of mildewed tobacco leaves based on machine vision and MSD descriptors

为实现醇化烟叶中霉变烟的自动在线精选,设计了基于机器视觉的霉变烟在线检测系统。该系统通过高速线阵CCD动态获取烟叶图像,采用MSD微结构描述算法提取烟叶图像颜色、纹理特征,基于神经网络集成分类算法,通过合格烟叶样本和霉烟样本的训练学习,实现霉变烟的在线检测识别。经过测试,该检测算法对霉烟图像样本的测度为0.918。在线检测试验结果显示,采用霉烟靶物单独过料时,机器视觉系统对霉烟的平均在线识别率在95%以上;将霉烟靶物与合格烟片混掺过料时,系统对霉烟的平均识别率在87%以上。研究结果表明,机器视觉方法用于醇化后烟叶中霉变烟的在线精选是可行的。

In order to realize automatic screening of mildew tobacco leaves, a moldy tobacco online detection system based on machine vision was developed. Tobacco leaf images were obtalned by high-speed linear CCD. MSD （Micro-structure descriptor） was used to extract image color, texture features. Neural network ensemble was adopted to generate individual learning algorithm, which realized the recognition of mildew tobacco leaves by tralning and learning. Recognition algorithm measure for mildew tobacco samples was up to 0.918. Results showed that recognition rate of target object was higher than 95% for mildew tobacco detection alone, while it was higher than 87% when mildew tobacco was mixed with normal tobacco leaves. It was indicated that the machine vision method is feasible for on-line selection of mildew tobacco leaves.