shksa/Multi-Digit-number-recognition-in-natural-images

Using the SVHN datset, a deep cnn model is trained in pytorch with a slightly different training routine and another straight forward cnn model in keras

actual321.ipynb has the training for the model on images with 1, 2, 3 numbers in it. The training routine is different from the usual method. A digit classifier gets backprop'ed only if it's the right classifier for the digit.

ex: for 2 digit image, only the first 2 digit classifier's are backprop'ed. 3rd, 4th, 5th are not backprop'ed.

saliency map.ipynb has the saliency visualization for the model trained in actual321.ipynb.

See SVHN folder for experiments on different models for the same task.(More interesing!).

model.py in this folder is a straight forward approach that achieves very high accuracy

Multi-Digit-number-recognition-in-natural-images

Acheived a test accuracy of 91%

Run test.py to check accuray

Test.ipynb is nice visualization of the model accuracy

The trained weights are in the file "phase_4.hdf5"

Used batch-normalization and dropout

Used max-norm constraint on weights instead of L2 regularization

He-normal initialization

5 classifiers for 5 digits

Absence of a digit is realized by using digit "10"