yogeshluthra / MultiDigit_detection_in_natural_scene

Multi Digit Detection Natural Scene

Following architecture was used (based on Goodfellow et al 2014 [1] and VGG16 [2] )

• Final Report on this project can be found in this repo
  • Final_Report.pdf
    • This report has links to the presentation video and demo video

Implementation worked well on most images tested.
Example Results

• All training steps could be seen in Train_FullyTrainVGG16_Dataset1_2xCrops_AugmentedWithFalseImages_withRandomRotations.ipynb

• The full environment, in which this project was developed could be found in:

  • requirements.txt

• Unfortunately Github doesn't host large files (>100MB). Hence trained model couldn't be uploaded.

  • But it can downloaded from this link
    • After download, this should be placed as ./trained_models/multi_digit_classifier_FullyTrainVGG16_Dataset1_2xCrops_AugmentedWithFalseImages_withRandomRotations

• To test your own images (after downloading the model)

  python run_v3.py [-h] -i IMAGE [-r ROTATION] [-md MINDIM_SIZE] [-n N_EXPANSIONS]

  optional arguments:  
  -h, --help            show this help message and exit  
  -i IMAGE, --image IMAGE  
                        Path to the image  
  -r ROTATION, --rotation ROTATION  
                        Rotation to be applied to image  
  -md MINDIM_SIZE, --minDim_size MINDIM_SIZE  
                        max pixel limit on minimum dimension of image (whether  
                        width or height)  
  -n N_EXPANSIONS, --n_expansions N_EXPANSIONS  
                        number of box size expansions to be applied  

References: [1] "Multi-Digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks" Goodfellow et al. 2014
[2] “Very Deep Convolutional Networks for Large-Scale Image Recognition”, Karen Simonyan, Andrew Zisserman 2015.
[3] Pre-trained VGG16 on ImageNet. https://keras.io/applications/#vgg16
[4] Batch Normalization, Sergey et al 2015
[5] Delving Deep into Rectifiers, He et al 2015
[6] Dropout, Srivastava et al.
[7] Street View House Numbers (SVHN) Dataset
[8] Google Street View Data Set
[9] On the importance of initialization and momentum in deep learning, Sutskevar 2013
[10] http://rodrigob.github.io/are_we_there_yet/build/classification_datasets_results.html#5356484e