CoordGate: Efficiently Computing Spatially-Varying Convolutions in Convolutional Neural Networks

Optical imaging systems are inherently limited in their resolution due to thepoint spread function (PSF), which applies a static, yet spatially-varying,convolution to the image. This degradation can be addressed via ConvolutionalNeural Networks (CNNs), particularly through deblurring techniques. However,current solutions face certain limitations in efficiently computingspatially-varying convolutions. In this paper we propose CoordGate, a novellightweight module that uses a multiplicative gate and a coordinate encodingnetwork to enable efficient computation of spatially-varying convolutions inCNNs. CoordGate allows for selective amplification or attenuation of filtersbased on their spatial position, effectively acting like a locally connectedneural network. The effectiveness of the CoordGate solution is demonstratedwithin the context of U-Nets and applied to the challenging problem of imagedeblurring. The experimental results show that CoordGate outperformsconventional approaches, offering a more robust and spatially aware solutionfor CNNs in various computer vision applications.