This project is a part of the Bill.com challenge.
Our motivation for the problem was to avoid using the OCR data completely by training a neural network to predict the receipt amount and date from the receipt image. We found that the date formatting was heterogeneous, so we calculated the number of days since 1/1/1970 and used that to train the network. From here, we used two densenet264 neural network models (one for the amount, the other for the date) to train the network. Afterwards, we took the output predictions per image and searched for them in the test_transactions.csv file.
Test results: Average difference between predicted and ground truth values: Amount: 0.1394 Date: 136.91
Log in or sign up for Devpost to join the conversation.