Research On Color Barcode Detection And Receipt Processing In Customs Applications

With the rapid development of international trade and cross-border e-commerce, more and more data sources need to be taken into consideration in customs tax man-agement and risk decision. Not limited in conventional structured data, information extraction and analysis from unstructured data is also required. As an important source of unstructured data, extracting useful information from document image carries great significance. This paper mainly focuses on two kinds of image document, that is, barcode image on commodity packaging and scanned documentation accompanying customs declaration.For barcode detection and recognition, this paper presents a fast color barcode detection method through cross identification, which considers color barcode as a grid of squares with 4 or 8 colors. A feature detector based on color differentiation is used to detect cross points, while thresholding and connected component analysis methods are performed to generating candidate barcode regions. Candidate regions containing more than a specific number of crosses while meeting certain shape conditions, are considered to be color barcode areas. On the self-built data set with about 600 color barcode photos, the method achieves high accuracy, simultaneously ensuring efficiency and robustness. For color barcode prototype proposed in this paper, we also present a size inferring method, which performs projective transformation first followed by edge extraction. Through analysis of distances between edges, we finally get barcode’s size and the corresponding matrix with specific color index in it. This method achieves good results for clear images with color barcode on a plane.For preprocessing of scanned documentation accompanying customs declaration, this paper presents a line detection method based on one-dimensional convolution.Firstly, we use a ring-shaped median filter to eliminate clutter noise. Then one-dimensional integrals along both horizontal and vertical directions is performed. After normaliza-tion and binarization of the integral image, connected component analysis is used to ex-tract a set of features, followed by SVM training and prediction. The method achieved high precision and recall on 55 real document images.