SVHN-Multi-Digit-Recognition

A implementation of Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks (http://arxiv.org/pdf/1312.6082.pdf)

tensorflow version

Graph

Accuracy

Loss

Result

Results on train, val and test dataset,

Inference of outside images

digit "10" means no digits

When length less than 5 the result confidence is quite high, but length is equal to 5, the result is usually wrong.

This is mainly because that the number of this kind of training images is quite small.

Requirments

• Tensorflow
• h5py
• Pillow

Setup

1. Clone the source code
2. Download SVHN Dataset format 1
3. Extract to data folder, which looks like this:

SVHN
        - data
            - extra
            - test
            - train

4）Pretrained Model can be download from my Baidu Cloud Drive and Password is crdf

Usage

1. Convert to TFRecords format

    python convert_to_tfrecords.py --data_dir your_data_dir

2. Train

    python train.py --data_dir your_data_dir --train_logdir ./logs_train

3. Evaluate

    python eval.py --data_dir your_data_dir --checkpoint_dir ./logs/train --eval_logdir ./logs/eval

4. Visualize

    tensorboard --logdir ./logs

5. Inference

    python inference.py --image image_dir --restore_checkpoint ckpt_dir/model.ckpt

pytorch version

To be continued...

Acknowledgement

• Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks

• In addition, I mainly reference potterhsu's implementation (Who implements tensorflow and pytorch version). I modify the model he used a little and achieve equivalent accuracy.