Machine Learning Intern : Mitek Systems Inc.

Poster, Slides

I spent Summer 2019 at Mitek Systems Inc. working with the document tooling team to improve document onboarding process and increase efficiency of modellers. The main projects that worked on were:

• Segmentation and four-corner detection of ID-documents
• Improved Image-Matcher Tool

Segmentation and four-corner detection of ID-documents

I designed and developed a novel algorithm based on U-Net architecture for ID-document segmentation in natural scenes. This was followed by drawing a convex-hull, k-means clustering and IoU optimization to detect four-corners of the document. The algorithm achieved an IoU of 0.98 for segmentation and median error of 7 pixels for four-corner detection.

Improved Image-Matcher Tool

I developed a complete pipeline incorporating automatic detection of ID-documents (described above) along with advanced compter vision algorithms for finding images of the same document type given a query image and a set of test images. The tool improved the legacy matcher by rejecting over 92% of False positives (FPs) while retaining over 80% of True Positives (TPs). I also designed an easy-to-use User Interface (UI) for the tool and successfully handed it over to the Document Modelling Team.