Python-Tensorflow Number Recognition

Neural Network trained to detect consistent-font numbers amongst large dataset complex images.

Related Tags...

PythonTensorFlowComputer VisionOpenCV

Screenshot Gallery...

Network Training Data

Class	# of Training Images
Zero	1368
One	588
Two	144
Three	109
Four	105
Five	184
Six	119
Seven	324
Eight	558
Nine	493

Network Structure

Layer #	Specific Type	Settings
0	Random Flip
1	Random Rotation
2	Random Zoom
3	Rescaling
4	2D Convolutional	16 filters, 3 kernel size Relu activation
5	2D Max Pooling
6	2D Convolutional	32 filters, 3 kernel size Relu activation
7	2D Max Pooling
8	2D Convolutional	64 filters, 3 kernel size Relu activation
9	2D Max Pooling
10	Dropout	Rate of 0.2
11		128 units Relu activation
12		10 units Relu activation

Training Stages

Initial Training:
Initially, the algorithm was applied to separate all number within the image (as determined by the strongest found contours).

Self-supported Training:
With a functional, yet inaccurate network running, a classification was run on all images in the set.
The results were then manually filtered, creating a significantly large set of accurate data.
The neural network training is run on this larger dataset, creating a significantly higher accuracy.

Algorithm

User provides a file-name (in the form of a number).
Image is opened, and preprocessed using a combination of Gaussian Blur, Canny Edge Detection, External Edge Detection
Take the four largest contours (most likely the numbers on the image)
For each contour, Blur, grayscale, and convert to Network Image required sizes
Neural Network run for each number, and value is concatenated and returned

Limitations

Correctly Identify all 4 numbers: 1294/1324 (97.734%)
Correctly Identify all 4 non-corrupted numbers: 1300/1324 (98.187%)
With the accuracy of the neural network, it appears that the error lies in incorrectly canny-edged images. Different threshvalues or a blurring method may provide an increase in accuracy