A multilayer perceptron that classifies digits (0–9) from images of 7-segment displays. After systematically testing 27 training runs across image sizes, hidden-layer widths, and preprocessing methods, the best configuration reached 99.79% test accuracy.

Best model

• Preprocessing: Sauvola thresholding combined with HOG (Histogram of Oriented Gradients) on 128×128 images
• Architecture: single hidden layer of 128 ReLU units, 10-way softmax output
• Training: Adam optimizer, categorical cross-entropy, 50 epochs, batch size 250
• Result: test accuracy 0.9979, test loss 0.0308

What the experiments showed

• Sauvola + HOG preprocessing at 128×128 dominated the leaderboard; the top nine runs all used it.
• Larger input resolution helped: 128×128 consistently beat 32×32 and 64×64.
• Raw grayscale without preprocessing topped out around 98%, while Sauvola-only at 128×128 was the weakest at 77.5%.