A reimbursement bill image segmentation method and system
An image segmentation and bill technology, applied in the field of computer vision, can solve the problems of perspective deformation, complex bill background, complicated collection methods, etc., to reduce the difficulty and improve the robustness of the algorithm.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Publication Date
- 2019-05-10
Smart Images

Figure 1 
Figure 2 
Figure 3
Abstract
Description
technical field
[0001] The invention relates to the field of computer vision, in particular to a method and system for cutting reimbursement bill images. Background technique
[0002] With the rise of the financial cloud in recent years, various bill reimbursements have begun to go to the cloud. The financial cloud managers the company's various reimbursement bill images, financial data, input and output items, and other information. One of the key steps in migrating bills to the cloud is that the bills need to be digitized, that is, the bills need to be photographed or scanned, and then divided into independent bills, and then the face information of the bills is identified, and finally the image and the recognized face information are stored in the financial database according to the financial logic. cloud.
[0003] The cutting and correction of the sticker image is a key link in the financial cloud bill processing step, which determines whether the bill can be recognized...
Examples
Embodiment
[0096] Step 1: Input the original image
[0097] Input the original image into the reimbursement receipt image segmentation system.
[0098] Step 2: Feature marker positioning
[0099] There are a variety of feature marks on the original image, including blue train tickets [1], red train tickets [2], printed oval stamps [3], two-dimensional codes [4], round stamps on the back cover [5], back cover Fangzhang [6] and other six categories. pass image 3 The fully convolutional neural network (FCN) shown is a UNet structure, down-sampled for feature extraction, then up-sampled to gradually fuse the feature information of each scale, and finally output the predicted feature score map, extracting 7 predictions Figure scoresmap, the value range of each scoresmap is between [0,1], the 0th scoresmap is the background of the original image. Then, for each scoresmap image, the pixel whose score is greater than the first set threshold (0.5) is considered to be a feature marker pixel, ...