STN-OCR

Detecting and recognizing text in natural scene images. This is still an open problem for research community and has many usge like image-based machine translation, autonomous cars or image/video indexing. The Algorithm consist of two stages which are 1.Text Dectection 2.Text Recognition stages. The Text Dection stage uses Resnet-Cifar version of Deep Residual Learning for Image Recognition ("https://arxiv.org/abs/1512.03385") and Spatial Transformer Network by Max Jaderberg ("https://arxiv.org/abs/1506.02025").The Text Detection and Recognition Stage again contains a variant Resnet-cifar version.The whole model is trained end-to-end.

Dependencies

• Python-3.x
• Tensorflow-2.3.1
• Opencv-4.x
• Numpy
• sklearn

Repository Description

• main.py script creates the whole model consiting of localisation network, Grid generator and sampler and Recognition network.
• stn_network.py script crates spatial transformer network, Grid genrator and bilinearsampler.
• resnet_stn.py script creates detection and recognition resnet network as proposed by the author.

Dataset

• The Street View House Numbers (SVHN) Dataset.[http://ufldl.stanford.edu/housenumbers/]
• Google FSNS dataset.[https://rrc.cvc.uab.es/?ch=6]

*Note-The wok is in progress and the repo will be updated frequently.

Description

• Update the runner code to take command line arguments instead of reading it from Json
• Replaced Segmentation model from 79 to 47.

Changes Made

• Updated the code to take CLI arguments.
• Removed dependency from Json file.
• Removed ankle and hip related code.
• Changed Segmentation model to 47.
• Updated Dockerfile.

Related Issues

Additional Notes

• The code output has dependency on mymako_osteophytes and mymako_bmd.

Merge Request Checklists

• Code follows project coding guidelines.
• Documentation reflects the changes made.
• I have already covered the unit testing.