To solve this test problem, I have used single-shot multi-object detection model (SSD -300) for reading the encoded text and identifying its location in the test image. SSD model is based on VGG-16 and the approach defined in this paper by Wei Liu.
Dataset used to train the model is Street View House Number dataset. (SVHN) Model is trained to detect digits from 0 to 8 but not 9. (Reason: Rotation augmentation is used and 9 is detected as 6).
Training set used = SVHN Training set +SVHN Extra Training Set
| Model SSD-300 | Training Set = 77% mAP | Test Set = 68% mAP approx |
|---|
I have used Python 3.6.5 and Tensorflow 1.12.
- Step 1 Run train.py file
The Program will take approx 45 min. to download SVHN dataset and start training. The hyper-parameter values used in the program are.
• Batch size = 32
• Learning rate
| Global Step | Learning rate |
|---|---|
| 0 – 2100 | 0.001 |
| 2101- 15000 | 0.0005 |
| 15001- 20000 | 0.0001 |
| 20001- further | 0.00001 |
• Momentum parameter for Momentum optimizer = 0.9
• L2 regularization factor = 0.0005
• Probability Threshold = 0.5
• Category to be classified by neural network = 9 (digits) + 1 (background) + 4 (location coordinate)=14
- Step 2 Run train_restore.py
In case train.py break, run train_restore.py. Changes to be done in every restore are.
• Update latest metadata file name to restore program to start training.
• And update Step parameter value to continue training.
- Step 3 Run detect.py
• To check results of training on Test set. Some more sample images from Test set can be found in folder More_Result_Images in repository.
- Step 4 Run detect_robot1.py
To get same result shown in Readme.md file above. Update latest metadata file name to read text from the image robot1.png.
Check Results:
• My latest trained model.chkt-final files can be downloaded from here.
• detect_robot1.py can be run directly to get same results.