Introduction
This project presents a convolutional neural network (CNN) model for image deblurring. The blur type is generated by randomly translating and overlaying slightly shifted original images. This type of blur simulates hand shake during long exposure photography.
Dataset and Preprocessing
The project uses the MNIST dataset, with images preprocessed as follows:
Normalized between 0 and 1 Split into train (80%) and validation (20%) Blurred images generated with random shift (using generator function)
Model
The model is a CNN with:
ResNet blocks to handle vanishing gradients ReLU activation for computational efficiency Batch normalization for stability and speed MSE loss function Adam optimizer with learning rate 0.0002 and beta_1 0.5
Training
The model was trained for 99 epochs on batches of 64 images.
Evaluation
The evaluation was conducted on 10000 blurred images from the validation set, repeating the experiment 10 times and calculating:
Mean Squared Error (MSE) MSE Standard Deviation
Results
Mean MSE: 0.0012 MSE Standard Deviation: 0.0001
Visualization
Some images are displayed to show the deblurring effect.