基于深度语义分割的重叠烟叶分级方法

Overlapping Tobacco Ranking Method Based on Deep Semantic Segmentation

目前,烟叶主要依靠技师手工去分级,耗时耗力,效率低下,现有的一些自动烟叶分级方法在实际应用中分级准确率偏低,对于重叠在一起的烟叶更难以给出较好的分级结果。为此,提出一种基于深度语义分割的重叠烟叶分级方法。利用语义分割网络DeepLabv3+将重叠的烟叶进行分割,提取分割后的单片烟叶的颜色、纹理、形状特征,采用F-score对提取的特征进行筛选,使用多个支持向量机集成学习的方法进行分级。实验得出在多个单片烟叶数据集上平均分级准确率达到71.23%,在两片烟叶重叠数据集上的实验达到48.49%的准确率。

Contemporarily,tobacco-leaf ranking is mostly dependent on technicians,which is time-consuming and inefficien-cy.A few existing automatical tobacco-leaf ranking methods cannot achieve satisfying accuracy in real applications,especially for overlapping leaves,which has been recognized as a much more difficult task.In this paper,a deep semantic segmentation-based overlapping tobacco-leaf ranking method is proposed.It employs the deep network DeepLabv3+to separate overlapping leaves.Then it identifies individual leaf by extracting the color,texture and shape features with an F-score feature selection scheme and training multiple support vector machines in an ensemble learning manner.Experimental results on multiple single-leaf datasets report an av-erage accuracy of 71.23%,and an accuracy of 48.49%on the overlapping tobacco-leaf datasets.