摘要
有效减少烟叶中的非烟物质,特别是减少在加工过程不慎混入的次生非烟物质一直是烟草工业中的挑战性问题。本文提出了一种基于MCFnet (Multiple Convolution Fusion U-NET)模型的高光谱非烟物质检测方法。它通过在U-NET模型的下采样段、中间段和上采样段分别使用中心差分卷积残差模块,快速傅立叶卷积残差模块和常规卷积层的特定混合,实现高质量的非烟杂物分割结果,完成非烟物质的定位与类型识别。模型的训练使用加权交叉熵和Dice损失函数的组合以保证模型经充分训练后能得到较好的性能。测试结果表明:对面积不小于10 mm × 10 mm的8类典型非烟杂物的检测,虚警率为1.12%,漏检率为0.45%,对不考虑杂物类别,即把所有非烟杂物作为一类,且包含了部分训练集中未出现的杂物类别的检测,虚警率为3%,漏检率为2.4%。与现有非烟杂物检测方法比较,性能有显著提升。此外,已在湖南中烟下属复烤厂完成了系统的实际部署与模型性能验证。
A challenging problem in tobacco industry is to efficiently decrease non-tobacco-leaf materials and especially the secondary produced ones during tobacco production. In this work, we present a non-tobacco-leaf materials detection method by using hyper-spectral images and the proposed Multiple Convolution Fusion U-NET. The method integrates center differential convoltion residule modual, fast Fourier transformation convolution residule modual and standard convolution into the U-NET model and respectively deploys each of them to the downsampling stage, center stage and upsampling stage of U-NET. To this end, the quality of impurity segmentation can be guaranteed, which further achieves accurate localiation and recogtion. We combine weighted cross entropy loss and Dice loss to train our model as we find it effectively and efficiently for sufficient training. The overall method achieved a false-alarm rate of 1.12% and a missed-detection rate of 0.45% for eight typical non-tobacco materials with the size less than 10 mm × 10 mm. When test in the case of not considering the types of non-tobacco materials, and a number of new non-tobacco materials being involved, the proposed method achieved 3% false-alarm rate and 2.4% missed-detection rate. A significant improvement is also found when compared with existing non-tobacco-leaf materials detection method. Moreover, system deployment and model performance validation have been finished in a tobacco factory of Chinese Tobacco Industry Hunan Co. Ltd.
出处
《图像与信号处理》
2024年第1期76-91,共16页
Journal of Image and Signal Processing
