Extraction Of User-entered Components From Filled-in Parcel Form Image Based On Image Registration

With the development of e-commerce, the number of online transactions has increased dramatically, which results in a large number of parcels needed to send and receive, sort and transport. Due to the huge number of parcels, automatic parcel sorting machines are in great demand. When a parcel is transporting in a sorting machine, its image is captured by a digital camera. Image processing and recognition technology is then used to extract way-bill form from the parcel image, and further identify the barcode, the recipient’s address, the recipient’s phone number and other information for automatic sorting. Extraction of user-entered components from filled-in parcel image is one of the technical issues.In this paper, we proposed a method to extract user-entered components from filled-in parcel form image based on the technology of image registration, for removing tables, LOGOs, other pre-printed words in the form. The processing is a preceding step for the downstream recognition of address, postcode, barcode, etc. Our main work includes:1. We propose a method which uses maximum clique to combine the geometric information of image patches with the similarity of SIFT descriptors to match control points. The proposed method is able to deal with the form images have different resolutions, brightness, directions, rotations, scales, etc., and obtain an accurate transformation model;2. We propose an image subtraction operation. Before the pixel subtraction of two images, the brightness of two images is adjusted to be consistent to adapt to different illumination. Because the registration is not completely accurate, we calculate the most appropriate grayscale value by searching a certain range of corresponding pixels during the pixel subtraction, for generating a self-adapting template image. The final result image is determined according to the template image and the input parcel image. Experiments show that this method achieves satisfactory results.The proposed method was tested on the SR1-PARCEL1 data set. SRI-PARCEL1 data set is consists of real parcel form images obtained by a parcel sorting machine of Chinese Post. Our method has achieved good results in the data set.