Cigarette Authentication by Using Deep Learning-Based Anomaly Detection of Adhesive Mark Images

Machine vision has demonstrated effective discrimination between genuine and counterfeit cigarettes. However, the constructed identification model is limited to only a few specifications of cigarette products, and its versatility is constrained. Additionally, due to challenges in obtaining fake cigarettes and the instability of their features, it is difficult to extract representative fake sample features. This can severely compromise the reliability of the identification method. To address these limitations, we propose a novel cigarette authentication method that employs deep learning-based anomaly detection of adhesive mark images. Our approach involves a self-designed image acquisition device to capture images of cigarette packet adhesive marks. We then remove the partial background of images to obtain adhesive mark image samples. Finally, we train the convolutional neural network classification model using only real samples, then generate the maximum predicted probability value based on the trained model for the sample to be detected and compare this value with a threshold to achieve authentication. Genuine cigarette packets produced by 10 types of packaging machines and counterfeit cigarette packets were used as objects, covering 37 cigarette specifications. Three classic neural network models were selected for experiments and 787 samples were identified. The results demonstrate that all models have authentication capabilities when using the mean value as the threshold, with InceptionV1 having a discrimination accuracy, sensitivity, and specificity of 94.41%, 94.22%, and 94.69%, respectively. And the time for all models to identify a single sample is around 10 ms. The proposed method is both simple and efficient, as it can construct a discrimination model without using any fake samples. Furthermore, the ability to a variety of cigarette specifications indicates a high practical application value.