Image_Restoration

Image restoration is one of the most important areas in imaging science. Model-based optimization methods and discriminative learning methods have been the two dominant strategies for solving various inverse problems in low-level vision. This paper aims to train different set of images on different models for denoising and deblurring. For denoising salt pepper and Gaussian Noise is introduced in the datasets, and for deblurring Gaussian Blur is used to train the model using CNN(Convolutional Neural Network) and Autoencoders on different datasets. This papers uses the metrics PSNR and SSIM. Experimental results show that Autoencoders in case of salt pepper denoising, in case of gaussian noise and CNN(Convolutional Neural Network) in case of Gaussian Deblurring gives better results.

---

Dataset

The link of the dataset used for denoising and deblurring is: https://drive.google.com/drive/folders/1rBxspGhADY_40jWQ5JZ9jtljWF_zryXd?usp=sharing

---

Conclusion

After running our dataset through different models, we saw that only few models shows the best result in each section. This section contains the final model for Denoising and Deblurring respectively.
We performed Denoising on two different types of noises, Salt & Pepper and Gaussian Noise and Deblurring on Gaussian Blur.

Denoising

Salt and Pepper Noise

For Salt and Pepper Noise, Autoencoders turned out to be the best fit. The dataset used contains 51 images. This dataset was created by inducing salt-pepper noise randomly all over. PSNR of 33.14 in case of Autoencoders was quite satisfactory as compared to the other models, and SSIM of 0.926 was also good. We got the desirable results of denoising Salt and Pepper Noise by the Autoencoders Model. The results for the same are shown in the figure 1.

                                                                  Figure 1: Salt and Pepper Noise on Autoencoders
The above graphs shows the Accuracy vs Epoch and Loss vs Epoch respectively for the Autoencoders Models. Below are the images before and after running the model. One can see the different between the noisy image, denoised image and the ground truth respectively.
                                                        
                                                        

Figure 2: Noisy Image, Denoised Image, Ground Truth respectively for the Autoencoders Model of Salt and Pepper Noise

Gaussian Noise

For Gaussian Noise, data set of 51 images was used and was run on two different models Convolutional Autoencoders and Convolutional Neural Networks. Both the models were run on the same dataset with the same noise for 100 epochs. The loss function giving the best results was mean squared error. For the gaussian noise convolutional neural networks model was finalised since it gave the best result. The metric finalised in this model was SSIM. The SSIM values for the noisy and predicted image came out to be 0.19 and 0.636 respectively. The results for the same are shown in the figure 3.

Figure 3: Gaussian Noise on CNN

The above graphs shows the Accuracy vs Epoch and Loss vs Epoch respectively for the CNN(Convolutional Neural Network) Model.

Figure 4 : Noisy, Ground Truth and Predicted(Denoised) images respectively for the CNN Model for Gaussian Noise.

Above are the images of the Noisy, Ground Truth and Predicted(Denoised) images using the CNN model for Gaussian Denoising of the images.

Deblurring

Gaussian Blur

For Gaussian Blur, CNN(Convolutional Neural Network) Model turned out to be the best fit. The dataset used contains 51 images and was created by inducing Gaussian Blur randomly all over. Validation PSNR of 34.73 and Training PSNR of 35.57 in case of CNN(Convolutional Neural Network) Model was quite satisfactory as compared to the other models. Hence, as we got fairly satisfactory results from our CNN(Convolutional Neural Network) Model, it is our desired Model for Gaussian Blur Deblurring. The results for the same are shown in the figure given below.

                                                                  Figure 5: Gaussian Deblurring using CNN Model

The above graphs show the ’Loss vs Epochs’ and ’PSNR vs Epochs’ respectively for CNN(Convolutional Neural Network) Model.

Figure 6: Input(Blurred), Ground Truth and Predicted(Deblurred) images respectively for the CNN Model for Gaussian Blur.

Above are the images of the Input(Blurred), Ground Truth and Predicted(Deblurred) images using the CNN model for Gaussian Deblurring of the images.