Description
This model is a convolutional neural network built with transfer learning using the pre-trained model 'VGG16.' The 'block5_conv1' layer is retrained, and a final dense layer with 128 neurons is added.
The model will be used as a preliminary step in solving Sudokus through linear programming. Model It is responsible for classifying the content of each sudoku cell:
- class_0: empty cell
- class_1: cell contains the number 1
- class_2: cell contains the number 2
- class_3: cell contains the number 3
- class_4: cell contains the number 4
- class_5: cell contains the number 5
- class_6: cell contains the number 6
- class_7: cell contains the number 7
- class_8: cell contains the number 8
- class_9: cell contains the number 9
The dataset is constructed with balanced classes using images from the famous "MNIST digits classification" dataset, as well as images of numbers written digitally.
Dataset schema
The image size it is 28x28 pixels. After applying data augmentation to the dataset, the total number of images is as follows:
- Training images: 5,600
- Validation images: 2,400
- Test images: 2,000
Test Accuracy: 0.9810
Other validations:
An initial validation is performed. It remains pending to increase the size of the validations to understand the reliability of the mode
The results of the inference are as follows:
