Shared-kernel Wavelet Neural Networks for Poisson Image Reconstruction
A new neural network with only 200 parameters reconstructs images in real-time from sparse Laplacian fields.
A team of researchers led by Yuanhao Gong has introduced a novel approach to image reconstruction that leverages the sparsity of Laplacian fields. Their paper, submitted to arXiv, demonstrates that the Laplacian field of an image is typically sparse and follows a stable distribution. This property allows for unique reconstruction of the original image by solving the Poisson equation with appropriate boundary conditions.
To solve this efficiently, the researchers developed a shared-kernel wavelet neural network. This network boasts less than 0.0002 million parameters (roughly 200), making it extremely compact for deployment on most devices, including edge hardware. Its linear computational complexity enables real-time reconstruction, and numerical experiments show it outperforms previous methods in accuracy. The method has broad applicability in image compression, low-light enhancement, and object tracking.
- The neural network uses only 0.0002M parameters (200 total), making it highly compact.
- It achieves linear computational complexity for real-time Poisson image reconstruction.
- The approach leverages the sparsity of Laplacian fields for higher accuracy than prior methods.
Why It Matters
Enables real-time, high-accuracy image reconstruction on resource-constrained devices, with broad applications in compression and low-light imaging.