IRIS: Smart Phone Aided Intelligent Reimbursement System Using Deep Learning

Financial reimbursement is considered as a cumbersome process of information extraction from actual invoices. Recently, scanner aided intelligent reimbursement methods using scanned images have been developed to reduce the manpower and process time but it lacks flexibility as well portability. Recently, smart phone is becoming more and more general in daily life and hence it is considered a natural way to develop smart phone aided intelligent reimbursement system (IRIS). However, it is difficult to effectively recognize the text information on the distorted and oblique invoices images in natural scene, which are took by smart phone, and the standard formatted output cannot be realized for the offset and misplaced text. In order to solve these problems, we propose an effective smart phone aided intelligent reimbursement method based on deep learning. First, we preprocess distorted images and then propose a hough transform accumulator (HTA) algorithm, which adds an accumulator based on the Hough transform to achieve tilt correction and image recovery operation of the distorted image. Second, in order to remove the redundant information on the invoice image in the natural scene, we apply the you only look once-version 3 (YOLOv3) algorithm to accurately locate, segment and intercept the key information areas on the invoice image. Third, we adopt the connectionist text proposal network (CTPN) to detect the import text information block areas in invoice images, and densely connected convolutional networks (DenseNets) to identify the detected text. The connectionist temporal classification (CTC) algorithm is added to the Densenets network to achieve alignment of the input and output formats of the text, accurate optical character recognition (OCR) is performed on the intercepted block area invoice image. Finally, we proposed a new algorithm Regular Matching and Recursive Segmentation (RMRS) based on recursive segmentation of regular matching, which performs standard formatted output on misaligned or offset information. The average accuracy of the recognition of optical characters in all block areas on the invoice is as high as 0.991 with a minimum of 0.962.