A Robust Approach to Detect Occlusions During Camera-Based Document Scanning

摘要

For the relationship between the company and the consumer to be safe, the consumer must prove his identity by sending a photo of his identification documents. These images present some occlusion challenges, as human fingers can occlude some relevant information, such as personal data and the wearer’s face. Current document imaging systems cannot deal with occlusion situations and return to the user feedback suggesting the correct way to hold the document. In this context, this work proposes an algorithm based on deep learning qualified for identifying if the document present in the image has or does not have occlusion. The proposed model can be used in real-time (video) applications, to quickly identify and discard a frame captured by the camera that contains occlusion, and guide the user on the correct way to hold a document, thus avoiding the capture of a document with compromised quality. The proposed approach comprises a segmentation module to locate the document, and a classification module to classify the document as with or without occlusion. The proposed model has a low computational cost and can be used in real-time on mobile or robotic applications.