aiSight

Banknote Denomination Recognition for the visually underprivileged

• Machine Learning model with CLR's, ReLU, Softmax, on Resnet-34 for detection, recognition of valid INR and USD.

• Model exported to PyTorch Script and deployed as a Mobile Application using Dart and as a Web Application using Render.

To convert ipython notebook to html or pdf use LaTeX or

jupyter nbconvert --to html "file_name".ipynb

Objective :

• Relay INR or USD banknote denomination if valid, retry otherwise

Inference

• Notched accuracy upto 90%, not bad for a multi-label ResNet-34 classification model

• Trivial INR Softmax out

• Trivial USD Softmax out

• F-Beta score - more weight on precision, less on recall

Goals :

• dataset collection upto a valid amount (>200)
• model accuracy upto 90%
• accuracy stays unchanged for unbalanced classes
• perspective warping
• Grad-CAM
• improve accuracy with custom-nn/resnet101 and larger dataset
• deploy on Android/iOS
• low bandwidth performance improvement
• minimize false negative probability (recall percentage)