abhigyan1120024/document_verification

Visual Recognition project for automating KYC Process in BANK

Document_verification

Medium Blog - https://medium.com/swlh/document-verification-for-kyc-with-ai-ocr-computer-vision-tool-3485d85d75f6

Visual_Recognition

DOCUMENT VALIDATION:

A RCNN based Image Classifier used to classify Aadhaar Card, Pan card and any other document. This model was trained on a dataset of Aadhaar Cards, PAN cards and Other documents like gas bills, voter ID cards, driving licence etc. collected from customer data. The model was trained over several variations of the images such as blurred or tilted images. The model has an accuracy of 94%.

Requirements

• keras
• Tensorflow
• OpenCv
• PIL
• Tesseract-OCR
• Google OCR

Training

• Due to privacy reasons I was given a very limited dataset, so I had to upsample this data to ensure that I could train my model for all possible scenarios. I used the ImageDataGenerator module from keras and added variations to the dataset like tilting or selective blurring and increased my dataset size for each class.

• Freezing the 4 layers allowed me to utilize only the convoluted layers and send the output to a custom fully connected neural network which would be adjusted for only specific images. Given that I had a limited dataset I had to include dropout in the fully connected network to prevent overfitting on the data.

• This process ensured that the model would understand only the specific document types required to be classified.

Evaluation

• The trained model was then evaluated on accuracy and through manual checking methods to ensure that the right cardsdocuments were being classified. The model has an accuracy of 94%.

• The operations team manually verified several images using the classifier to ensure that the right documents were classified.

Impact of the project

• This project was packaged into a ready to use library which helped save more 50 hours of manual customer document verification. This helped the operations team to focus on other high priority tasks.

Future Improvements

• The model can be trained on fake documents and this can help in detecting fraudulent documents.