基于深度神经网络的卷烟燃烧锥落头检测和分类识别

Detection and classification-recognition of combustion cone-falling based on deep neural network

为提高卷烟质量评价过程中燃烧锥落头倾向的判断效率和准确性,构建了卷烟燃烧锥落头检测和分类识别模型。测试了100种市售卷烟的燃烧锥落头倾向,对收集的卷烟燃烧锥图像进行缺陷剔除、分类标注以及特征增强或减弱等处理后分别建立落头检测数据集和落头分类识别数据集;选取ResNet、MobileNet和ViT(Vision Transformer)作为主干网络建立3个深度神经网络模型,利用燃烧锥落头图像对模型进行训练,并采用精确率(P)、召回率(R)、f1分数以及模型容量对模型表现进行评估。结果表明:①3个模型均能较好地进行落头检测(模型损失值L<0.05,P>99%,R>99%),但ViT模型在落头分类识别测试集上的损失值一直未能收敛,不适用于落头分类识别(P=86.05%,R=76.04%);②基于f1分数和模型容量优选出的MobileNet模型具有准确度高、运算速度快等优势,对落头检测和落头分类识别的平均精确率分别为99.64%和89.50%。该方法可为研究卷烟表观燃烧性能提供支持。

To better judge combustion cone-falling propensity of cigarette,the detection and classification-recognition models of combustion cone-falling were established.The testing of combustion cone-falling propensity was conducted with 100 kinds of commercial cigarettes,their images of combustion cones were processed via defect rejection,classification annotation and feature enhancing/weakening.With the processed images,a dataset of cone-falling detection and a dataset of cone-falling classification-recognition were set up separately.Three deep neural network models were established by using ResNet,MobileNet and ViT(Vision Transformer)as backbone networks and were trained with the images of combustion cone-falling.The performance of the models was evaluated with precision(P),recall rate(R),f1 score and model capacity.The results showed that:1)The three models performed well on cone-falling detection(model lose value L<0.05,P>99%,R>99%);however,the ViT model was not suitable for cone-falling classification-recognition due to its loss value in the dataset of classification-recognition failed to converge(P=86.05%,R=76.04%).2)The MobileNet model selected based on f1 score and model capacity had the advantages of higher accuracy and operational speed,its average precision for cone-falling detection and classification-recognition was 99.64%and 89.50%respectively.This method provides a support for studying the apparent combustion property of cigarette.