Document Cleaner is a deep convolutional autoencoder model for generating a clean form of dirty documents.
Autoencoders can be used for ignoring the noise from image data. Here is a work for it.
Model trained for 300 epochs and its last loss is 0.0030. Loss function is "mean squared error" and optimizer is a "adamax" with 0.001 learning rate. Last layer's activation function is a custom function which maps values between -0.5 and 0.5 and other layers' activation functions are linear function.
Other solutions to improve model:
-Getting more data by editing train images with rotating, transforming etc.
-Making experiments on more optimizer and more hyperparameters
-Trying more neuron counts on layers
Here is the loss plot of model:
Cleaned Dirty Document examples:








