Positional Encoding Image Prior.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep Image Prior (DIP) utilizes Convolutional Neural Networks (CNNs) to reconstruct images by leveraging the network's inherent image prior.
• The effectiveness of DIP is often attributed to the implicit image representations learned by CNNs.

Purpose of the Study:

• To re-examine the Deep Image Prior (DIP) framework through the lens of neural implicit representations.
• To introduce a new method, Positional Encoding Image Prior (PIP), that utilizes Fourier Features (Positional Encoding) instead of random latent spaces.

Main Methods:

• Replaced the random latent input in the DIP framework with Fourier Features (Positional Encoding).
• Empirically demonstrated that convolutional layers in DIP can be substituted with pixel-level MLPs due to Fourier feature properties.
• Proved the equivalence of these methods for linear networks.

Main Results:

• Positional Encoding Image Prior (PIP) achieves performance comparable to DIP on various image reconstruction tasks.
• PIP significantly reduces the number of parameters required compared to traditional DIP methods.
• PIP demonstrates effective and stable extension to video reconstruction tasks, overcoming limitations of prior methods.

Conclusions:

• Fourier Features provide a powerful alternative to random latent spaces in image prior frameworks.
• PIP offers a more parameter-efficient and versatile approach to image and video reconstruction.
• The neural implicit representation perspective offers new avenues for developing advanced generative models.