This is a CNN for image clasification of 4 dogs using a set of 4000 pictures
This code is divided in 4 stages:
- Image preprocessing
- Creation of numpy datasets
- Model definition, testing and fitting using Keras
- GUI creation to implement model using tkinter
This part rotates and resizes the images to a suitable image size to be computed swiftly. 300x225 px is the size selected and all images are rotated to be in a horizontal view. This part only uses PIL and os modules.
This section creates train and test sets in .npz format files to store the pictures and to prevent recalculation of datasets.
I created the following architecture:
| Layer | (type) | Output Shape | Param |
|---|---|---|---|
| conv2d | (Conv2D) | (None, 215, 290, 64) | 23296 |
| max_pooling2d | (MaxPooling2D) | (None, 53, 72, 64) | 0 |
| conv2d_1 | (Conv2D) | (None, 47, 66, 48) | 150576 |
| max_pooling2d_1 | (MaxPooling2 | (None, 15, 22, 48) | 0 |
| conv2d_2 | (Conv2D) | (None, 13, 20, 32) | 13856 |
| max_pooling2d_2 | MaxPooling2 | (None, 6, 10, 32) | 0 |
| flatten | (Flatten) | (None, 1920) | 0 |
| dropout | (Dropout) | (None, 1920) | 0 |
| dense | (Dense) | (None, 64) | 122944 |
| dense_1 | (Dense) | (None, 64) | 4160 |
| dense_2 | (Dense) | (None, 64) | 4160 |
| dense_3 | (Dense) | (None, 4) | 260 |
Total params: 319,252
Trainable params: 319,252
Non-trainable params: 0
I also used Adam optimizer with the following parameters
(which could be tuned and improved):
opt = keras.optimizers.Adam(learning_rate=0.0013, beta_1=0.87, beta_2=0.9, epsilon=1.9e-07)
It achieved from 97% to 99% accuracy. Other architectures (more sophisticated) were not tested due to hardware limitations.
To show the results I created a simple GUI using tkinter
