Classification of invoice images by using convolutional neural networks and text similarity

Today, all companies pay for the expenses made on behalf of the company for various reasons and the invoices for the payment are kept for accounting purposes. As the company grows, the number of personnel increases and the number of supplier companies that the company works with increases in parallel. With the increase in the number of personnel and supplier companies, the amount of expenditure on behalf of the company also increases. Depending on the increasing amount of expenditure, the number of invoices for expenditures also reaches large numbers. In order to access the information on the invoice images later, the information on the image must be stored. Storage of this information is also necessary for legal reasons. Invoices that are physically sent to the accounting units are transferred to the digital environment manually by the accounting support personnel. Since very large companies will have much more invoices, much more personnel power will be spent during the transfer of invoices to the digital environment. In cases where the staff cannot transfer huge amounts of invoices, more accounting personnel are recruited. In addition, when too many invoices are entered in a limited time, the amount of error made during entry also increases and additional time and effort is consumed for corrections. The purpose of this thesis study is to automate the manual invoice entry processes performed by the personnel working within the company. Accordingly, an architecture based on convolutional neural networks, which is one of the deep learning methods, was established and invoices were classified according to their templates. The classification process is made according to the location of the invoice fields on the invoice and the template structure. After the invoice is classified according to the template, OCR is applied to the invoice whose template has been determined, and the characters on the invoice are automatically read and transferred to the natural language processing unit. Natural language processing unit was used to determine which invoice field belongs to the data taken from the invoice image and to determine its position on the invoice. With the new method that requires minimum personnel power, the number of possible mistakes that the personnel may make during entry will be minimized, and the invoice entry process will become much easier and safer.