LogicIR: Logic Gate Networks for Image Restoration

Overview

LogicIR introduces a new paradigm for image restoration by replacing standard arithmetic-based convolutional layers with logic gate networks (LGNs). While traditional deep neural networks (DNNs) rely on floating-point multiplications and additions, LogicIR performs inference using only fundamental logic operations like NAND and XOR. This approach is designed to minimize computational overhead, making it highly suitable for resource-constrained hardware such as FPGAs and ASICs.

Architecture and Methodology

LogicIR adopts a UNet-inspired encoder-decoder structure to handle hierarchical feature representation, which is critical for tasks like denoising and deblocking. To overcome the inherent limitations of logic-based computation, the authors introduce three key innovations:

• Differentiable Bit Decoding: Instead of forcing the network to output raw pixel values directly, the model predicts multi-channel binary maps that are transformed into a residual image via a learnable scaling mechanism. This improves training stability and gradient flow.
• Index Shuffling: To prevent the isolation of information within fixed logic groups, the authors implement an index shuffling mechanism that reorders input indices across logic tree kernels, facilitating better information propagation.
• Rotational Ensemble: To compensate for the limited receptive field and fixed connectivity of logic gates, the model uses a rotational ensemble strategy during inference, averaging outputs from multiple rotated versions of the input to improve restoration quality.

Performance and Significance

Experimental results demonstrate that LogicIR achieves strong performance on standard benchmarks (e.g., denoising and deblocking) while maintaining a significantly lower binary operation count (BOPs) than existing lightweight models. By eliminating the need for floating-point arithmetic, LogicIR provides a viable path toward deploying high-quality image restoration models on low-power, embedded devices where traditional CNNs are often too computationally expensive.