ICE: Implicit Coordinate Encoder for Multiple Image Neural Representation.

Area of Science:

• Computer Vision
• Computer Graphics
• Deep Learning

Background:

• Implicit Neural Representations (INRs) offer advantages over discrete methods for representing signals like images and shapes.
• Scaling INRs to image collections is challenging due to rapidly growing parameter requirements.
• Existing INR methods primarily rely on Multi-Layer Perceptrons (MLPs).

Purpose of the Study:

• To propose a fully implicit approach for INR that drastically reduces model size for multiple image representation tasks.
• To introduce the Implicit Coordinate Encoder (ICE) for efficient scaling of INRs with image number by learning a common feature space.
• To demonstrate the applicability of the method to both image collections and large (gigapixel) images.

Main Methods:

• Developed a fully implicit INR approach using an auto-encoder architecture with a single ICE (encoder) and multiple MLPs (decoders).
• Employed a multi-task learning strategy for joint training of the ICE and MLPs.
• Implemented ICE as a one-dimensional convolutional encoder, integrating convolutional blocks into INR networks for the first time.
• Utilized a "divide-and-conquer" strategy for handling large images.

Main Results:

• Achieved significant reduction in MLP model size for multiple image representation tasks.
• Demonstrated effective scaling of INRs for image collections by learning a common feature space via ICE.
• Showcased the method's validity for large single images using a "divide-and-conquer" approach.
• Obtained better quality than previous fully-implicit methods with up to 50% fewer parameters on the Kodak dataset and a large Pluto image.

Conclusions:

• The proposed ICE method offers a scalable and efficient solution for applying INRs to image collections and large images.
• Integrating convolutional blocks within INR networks via ICE improves performance and reduces parameter count.
• This work pioneers the use of convolutional components in INR architectures, advancing the field of implicit neural representations.