The Fifth Elephant 2019

Identifying a credible signature match based on a base signature of a person is an age old problem. Despite recent automation and advances in this field using image recognition, a lot remains to be explored. We have developed an intelligent framework which can automatically detect a forged signature even if it is highly skilled, based on the developed feature embeddings and the corresponding algorithm. Siamese Triplet Convolution Neural Network is used to generate the feature embeddings for the signature images followed by a generalized Logistic Regression model to detect forgery. On the widely used SigComp dataset, our system achieves an accuracy of 96% in detecting forged signatures. Once the model is trained, it requires just one base image to determine whether another signature image is genuine or fraudulent with one shot learning. This algorithmic framework can be used in multiple commercial settings. One such example is testing customer or employee signatures on documents against a corresponding base signature saved beforehand.

Outline

The differences between the images of a genuine signature and its skilled forgery is at times very minute and is challenging to detect by even a trained eye. Deep Triplet loss function is a very powerful loss function used in the industry for face recognition. We have created our custom triplet model architecture with modified MobileNet CNN and dense layers with triplet loss function. Based on this loss, the image embeddings are created in such a way that the dissimilarity between the anchor image and positive image must be low and the dissimilarity between the anchor image and the negative image must be high for every triplet. This kind of architecture ensures that even small differences in signatures can be captured in order to flag a skilled forgery effectively.

The talk will cover the following:
• Introduction to the business objective of Signature Fraud Detection and the algorithmic challenges associated with Signature Fraud.
• A detailed domain background on various types of Forgery and their complexities.
• Description of the various Dataset sources and the complexities in data collection and forming the training set with image triplets.
• Introduction to Siamese Triplet CNN Architecture and one-shot learning applications in fraud detection.
• Transfer Learning and finetuning the siamese triplet model using MobileNet architecture and its benefits.
• Logistic loss function implementation with the above architecture to form a robust method in detecting fraud signature.
• Final Framework architecture of the entire methodology and the accuracy attained using the same.
• Reference and Conclusion.

Requirements

Deep Learning,Computer Vision,Machine Learning, Image Processing.

Speaker bio

Ojaswini Chhabra received M.S. (Quantitative Economics) from Indian Statistical Institute, Delhi and B.A. (Hons.) Economics from University of Delhi in 2016 and 2014 respectively. She currently works as Senior Statistical Analyst in Walmart Labs, Bangalore. She has worked on several classification and optimization-based problems and filed patents for the same. Her research interests include natural language processing and computer vision.
Email: ojaswinichhabra@gmail.com
Address: G-504, Tower 6, Adarsh Palm Retreat, Bellandur, Bengalore - 560103 Contact: +91-8285671581

Souradip Chakraborty received M.S. (Quality Management Science) from Indian Statis- tical Institute, Bangalore and B.E. (Engg.) Electronics and Instrumentation from Jadavpur University in 2016 and 2014 respectively. He is currently working as a Statistical Analyst in Walmart Labs, Bangalore. His current field of research interest lies in representation learning, mixed-space representations, NLP, Vision, and he has several patents filed (as a part of Walmart Labs) in the field of AI and Machine Learning applications.
Email: souradip24@gmail.com
Address: 603, Vasudeva Residency, 4th Cross Rd, Bansawadi, Bangalore - 560043 Contact: +91-9038790361

Slides

https://drive.google.com/file/d/14VPShWIH36VtSFKeyTIK-eUhr-klKwYf/view