Image-Denoising-Using-Encoder-Decoder-based-model

Understanding and Implementing the Image Reconstruction and Image denoising techniques in Computer Vision Area

Image Reconstruction

The autoencoder based model is used to reconstruct the image from its latent representation. The model is trained on the standard Fashion MNIST dataset with image size of 28x28 and one channel.

------------------> training image sample

Each training image has 28x28 = 784 features. The model will use the latent dimension of 64 to compress the image representation and using decoder block it will recreate the image similar to original image with minimum reconstruction error.

---------> Autoencoder model for image reconstruction

---

For Training, following scipt can be run:

python train_autoencoder.py --learning_rate 0.001 --train

---------------> training MSE Loss graph

After training the model, it can be used to reconstruct the input image like below.

-----------------> Image reconstructed using model

---

Image Denoising

The Encoder-Decoder based model will be used to capture the structural and local features of the image and construct image without noise to give denoised image. The model is trained on the standard Fashion MNIST dataset with image size of 28x28 and one channel. The normal distributed noise with noise_factor of 0.2 is added to each input image and model will try to learn to construct the image without noise.

------------------------> Input Image with noise

The model uses convolution and deconvolution layers to downsample and then upsample to construct output image with same dimension

----------------------> Image Denoising model summary

---

For Training, following scipt can be run:

python train_denoising.py --learning_rate 0.001 --train

After training the model, it can be used to denoise the input image like below.

----------------> Training loss graph

---------------> Image denoised by model

Use Cases: