Biliteral and Wavelet Denoising Filter 

Using Python, we applied various types of noise—salt and pepper, speckle, and Gaussian—to four images to explore denoising techniques. Initially, we employed classical methods such as wavelet denoising and bilateral filtering, implemented through Python libraries, and manually processed and saved the results. 

The two types of denoising we focused on were wavelet denoising and bilateral filtering. 

Bilateral filtering: 

• Applies a weighted average of nearby pixels based on proximity and intensity of color, but preserves edges where the intensity varies sharply.

• Smoothes areas with similar intensities while preserving sharp edges. 

• Works well for images where edge preservation is important but detail preservation is not

• Struggles with images with precise features, such as tree bark, or preserving fine details on a face.

Wavelet denoising:

• Image is decomposed into wavelet coefficients.

• Large features in the image are captured as low frequency (approximation coefficients). Small details are captured as high frequency (detail coefficients). Noise generally affects the high frequency components of an image.

• Uses thresholding to remove the noise from the high frequency.  For example, if a detail is below a certain threshold, it is removed.

• We expect wavelet denoising to remove more noise than Bilateral Filtering

As you can see from the images above, both bilaterally denoised images and wavelet denoised images have their flaws and strengths. The bilaterally denoised images have a lot of noise leftover and lose their color, but the edges are crisp. The wavelet denoised images have a bit less noise, but are blurrier than the bilaterally denoised images.

We found these techniques insufficient for significant noise reduction. As a result, we turned to MPRNet, a pre-trained model that leverages diffusion processes and neural networks to identify and enhance the underlying features of images, resulting in superior denoising performance. This model excels at reducing noise even in heavily corrupted images and can effectively handle images with multiple types of noise. 

Despite the improved noise reduction, a notable issue is the distortion of color in the output images. This color distortion occurs because the neural network's training process might not fully preserve the color information while focusing on denoising, leading to a compromise between noise removal and color fidelity. Additionally, the model's emphasis on structural features can sometimes result in altered color tones in the reconstructed images.

The pre-trained model leverages diffusion processes and neural networks to effectively identify underlying features in images, addressing challenges in noise reduction. The architecture integrates an encoder-decoder framework, which is then refined through a branch specifically designed to supervise local information. This branch employs reweighting mechanisms to prioritize local features based on specific areas of the image, ensuring a more context-aware output. Additionally, the model incorporates three weighted attention modules that strategically analyze and enhance these local features, resulting in improved image quality. Hyperparameter tuning was conducted to achieve optimal performance, and the model demonstrated its efficacy through tests on multiple noisy images, showcasing its capability to reconstruct images with reduced noise and improved feature retention.

Based on the results above, we can see how successful our denoising algorithms were when applied to these three filters. For Gaussian, we see that there is no noise remaining and the images are nearly perfect from the original. For speckle, we once again have no noise remaining but there images are a bit sharper, and a little harder to tell what the original photo is such as the third image. Finally, Salt and Pepper model performed worse than the three noises that we applied. We can see that the images are significantly harder to identify. In all, we are proud of the results that we have yielded but do want to improve on the algorithm and also try to identify why the color of the images changes.