Tobacco Foreign Object Detection using Deep Convolutinal Features

Accurately locating foreign object in tobacco raw materials is of great significance for the quality assurance of cigarette products. To handle this problem of foreign object detection in tobacco leaves, this paper proposes an effective detection method based on deep convolutional neural networks. The proposed method employs the widely-used U-Net encoder as a feature extractor to capture the features of tobacco images, which can overcome the problem that traditional feature descriptors lack adaptability in representational ability. Meanwhile, in order to further enhance the robustness of similarity measurement, we adopt a robust measurement method, which calculates the difference values of the deep features of each pixel based on feature blocks. The foreign objects in tobacco raw materials can be accurately located by applying an adaptive threshold on the difference map. The experimental comparison results verify the effectiveness and superiority of the method proposed compared to existing methods.