Using Deep ConvNet for Robust 1D Barcode Detection

Barcode has been widely adopted in many aspects, it is the unique identification and contains important information of goods. Regular barcode scanning device usually requires human being ¡¯s aids and is not suitable for multiple barcode scanning, especially in a complex background. In this paper, a cascaded strategy is proposed for accurate detection of 1D barcode with deep convolutional neural network. The work contains three parts: Firstly, a faster Region based Convolutional Neural Net (Faster R-CNN) framework is used to train a barcode detection model. Secondly, a powerful lo-level detector called Maximally Stable Extremal Regions (MSERs) is developed to eliminate the background noisy and detect the direction of the barcode. Thirdly, a postprocessing with like bilateral filter, called Adaptive Manifold (AM) filter, is applied when the image is blurred. We have carried out experiments on both Muenster Barcode Database and ArTe-Lab Barcode Database and compared with the previous barcode detection methods, the result shows that our method not only can get a higher barcode detection rate but also more robustness.