Classification of Cigarette Types Using Computer Vision: An Analysis of Smoke Aggregation Features

Abstract

The aggregation of cigarette smoke is one of the key indicators when consumers taste cigarettes. The purpose of this study is to explore the characteristics of smoke images through computer vision technology, and distinguish different brands of cigarettes by using the aggregation characteristics of smoke, hoping to provide preliminary support for automatic cigarette identification. Firstly, we constructed a 3D model experimental platform based on the human upper respiratory tract to create the cigarette smoke image dataset SmogAg. This data set contains sequence image data of smoke movement patterns of 51 types of cigarettes, including thick, medium and thin cigarettes. Based on this data set, we applied the traditional vision algorithm to extract the diffusion width, sedimentation concentration and diffusion distance of smoke, and used the KNN model for training, and initially reached an accuracy of 69.8%. To further improve performance, we have adopted a new network architecture. We first adopted SmokeSeqNet model for experiments, which combined residual CNN as a feature extractor and captured the dynamic sequence characteristics of smoke through the LSTM layer. This method, combined with time series analysis, not only improved the accuracy, but also gave us insight into the time-dependent behavior of the smoke, ultimately achieving an accuracy of 81.6%. In order to further improve the performance, we build SmokeSeqNetV2 model to continue the experiment, which introduces Swin-Transformer feature extraction module and attention mechanism, and the final accuracy rate reaches 88.9%. Through computer vision technology, we began to explore the image characteristics of smoke, hoping to provide a new idea for distinguishing different brands of cigarettes by the aggregation characteristics of smoke.

Keywords

• Smoke aggregation
• computer vision
• residual CNN+LSTM
• Swin-Transformer
• attention mechanism