RedCap: residual encoder-decoder capsule network for holographic image reconstruction.

Area of Science:

• Deep Learning
• Computational Imaging
• Optics

Background:

• Convolutional Neural Networks (CNNs) can suffer from information loss during pooling operations.
• Capsule Networks (CapsNets) offer an advanced deep learning approach to mitigate these limitations.
• Digital holographic reconstruction is a computationally intensive process.

Purpose of the Study:

• To explore the application of capsule networks for the first time in digital holographic reconstruction.
• To develop an efficient and accurate deep learning model for holographic data processing.
• To address the computational and memory constraints of existing reconstruction methods.

Main Methods:

• Development of a residual encoder-decoder capsule network (RedCap).
• Implementation of a novel windowed spatial dynamic routing algorithm.
• Integration of residual capsule blocks, extending the concept of residual blocks.

Main Results:

• RedCap demonstrated superior experimental results in digital holographic reconstruction compared to CNN-based methods.
• The proposed RedCap model achieved a 75% reduction in the number of parameters.
• Enhanced data processing efficiency and reduced memory storage requirements were observed.

Conclusions:

• RedCap offers a more efficient and effective solution for digital holographic reconstruction.
• The model's reduced parameter count and improved performance make it suitable for resource-limited applications.
• Capsule networks represent a promising advancement for holographic imaging and related computational tasks.