Evaluation of a classification algorithm for identifying filter vent blocking
Zwierzchowski, Benjamin A.
MetadataShow full item record
Significant inter-individual variation in cigarette smoke exposure from 'low tar' cigarettes can results from different styles of smoking cigarettes, such as puffing patterns and blocking of filter vents, collectively known as "compensation". A new method for classifying filter vent blocking behavior of smokers using ventilated cigarettes will be evaluated. The method, developed with JAVA and JavaFX imaging analysis technology, extracts a feature set of rotationally invariant color mean statistics from the digital image of a smoked filter. These features are used with a support vector machine (SVM) and logistic regression model, to train a classifier. The classifier uses the model to assess blocking classification scores to images of smoked filters of unknown filter vent blocking status, at the levels of unblocked, partially blocked, and fully blocked. Experimental sets using machine-smoked filters demonstrate significant improvement in classifying vent-blocking behavior against the state of the art. Two standards models were generated using images of Kentucky Reference 1R5F and 3R4F cigarettes. These standard sets were generated using total puff volumes ranging from 70 to 700 mL, at 0%, 50%, or 100% blocking conditions using cellophane tape. Our standards models yielded validation accuracies of greater than 95% versus the 'ground truth' condition. Data are analyzed on the classification of 'unknown' samples obtained from ITC human smokers. These results further pave the way for a low-cost, non-invasive digital imaging assay to help us to better infer smoking topography and yield a better understanding of smoking behavior and carcinogen exposure.