基于稀疏自编码器的烟叶成熟度分类.pdf

烟草科技/ Tobacco Science Technology 设备与仪器/Equipment Apparatus 2014 年第9 期（总第326 期） 基于稀疏自编码器的烟叶成熟度分类 王 杰，贾育衡，赵 昕 郑州大学电气工程学院，郑州市科学大道100号 450001 摘要：为降低人工成本，提出了一种基于稀疏自编码器的烟叶成熟度分类算法，从计算机视觉角度自动 识别烟叶的成熟度。首先对烟叶数字图像进行去除背景、归一化等预处理操作；其次从无监督学习算法 入手，利用稀疏自编码器构建特征学习网络；然后使用部分联通网络进行特征扩展，解决烟叶数字图像 像素过大、计算耗时长等问题；最后使用Softmax 回归对学习到的特征进行分类。结果表明：将稀疏自编 码器应用于烟叶成熟度分类，测试数据分类准确率可达98.63% ，优于其他分类器的分类精度。该算法直 接从像素层面提取所需要的特征，不需要人为选定设计特征，为提高烟叶成熟度分类效率提供了帮助。 关键词：烟叶；成熟度；稀疏自编码器；计算机视觉；特征学习；Softmax 回归 中图分类号：S572 文献标识码：A 文章编号：1002-0861 （2014 ）09-0018-05 Tobacco Leaf Maturity Classification Based on Sparse Auto-encoder WANG Jie, JIA Yuheng, and ZHAO Xin School of Electrical Engineering, Zhengzhou University, Zhengzhou 450001, China Abstract: For reducing labor cost, a leaf maturity classifying algorithm based on sparse auto-encoder was proposed, wherein leaf maturity was automatically identified by computer vision. The digital image of tobacco leaf was pretreated via background removing and normalizing at first; next, a feature learning network was established by means of a sparse auto-encoder on the basis of unsupervised learning algorithm; then, the features were expanded via a partially connected network as a solution to tobacco leaf image processing; finally, Softemax regression was used to classify the learned features. The proposed algorithm extracted the required features from pixel level directly without purposely selecting design features. The results of simulation showed that when the sparse auto-encoder was applied to tobacco leaf maturity classification, its accuracy of data classification reached 98.63%, which was higher than that of other classifiers. Keywords: Tobacco leaf; Maturity; Sparse auto-encoder; Computer vision; Feature learning; Softmax regression ［2 ］ 成熟采收是优质烟叶生产的关键技术之一，准确掌 影响经济效益 。烟叶成熟度的判断是一个综合评价 握田间鲜烟叶的成熟