Files: create_samples_no.py, cnn.ipynb
result:
Now all left is train then, number recognizing system is good to go.
To make Dataset use your haar_cascade on previous dataset (also resize images to a standard like 45 x 55) and manually segregate the result
img = cv2.imread('samples/'+<fileName>,cv2.IMREAD_GRAYSCALE)
numbers = num_cascade.detectMultiScale(img)
try:
x,y,w,h = numbers[0]
roi = img[y:y+h,x:x+w]
roi = cv2.resize(roi,(45,55))
cv2.imwrite('samples_no/'+<fileName>,roi)create_samples_no.py
Use create_samples_no.py for the same.
I have uploaded the number dataset for any other query.
Unzip samples_no.zip for my dataset!
For training CNN,I have made model and trained model in python (or google colab) and then converted it json for tf.js compatibility, you can also make model,train it in js/node/browser using tf.js and save it directly to json format.
CNN Model is quite simple just use 16 filters and then one pooling layer and dense it by a factor of 8 times and BOOM!!, CNN yield great results.
| Layer (type) | Output Shape | Param # |
|---|---|---|
| conv2d_9 (Conv2D) | (None, 53, 43, 16) | 160 |
| max_pooling2d_9 (MaxPooling2D) | (None, 13, 10, 16) | 0 |
| flatten_9 (Flatten) | (None, 2080) | 0 |
| dense_23 (Dense) | (None, 260) | 541060 |
| activation_19 (Activation) | (None, 260) | 0 |
| dropout_6 (Dropout) | (None, 260) | 0 |
| dense_24 (Dense) | (None, 10) | 2610 |
| activation_20 (Activation) | (None, 10) | 0 |
Total params: 543,830
Trainable params: 543,830
Non-trainable params: 0and finally save the model.
I have first saved the model in h5 format and then converted it to JSON. (as shown in last Part)
model.save('cnn/model.h5')but you can directly save it into JSON, read this for further insight.
import tensorflowjs as tfjs
tfjs.converters.save_keras_model(model, 'cnn')Also Check cnn.ipynb for Code.
Next Part