Abstract
Many image processing networks apply a single set of static convolutional kernels across the entire input image, which is sub-optimal for natural images, as they often consist of heterogeneous visual patterns. Recent work in classification, segmentation, and image restoration has demonstrated that dynamic kernels outperform static kernels at modeling local image statistics. However, these works often adopt per-pixel convolution kernels, which introduce high memory and computation costs. To achieve spatial-varying processing without significant overhead, we present \textbf{Malle}able \textbf{Conv}olution (\textbf{MalleConv}), as an efficient variant of dynamic convolution. The weights of \ours are dynamically produced by an efficient predictor network capable of generating content-dependent outputs at specific spatial locations. Unlike previous works, \ours generates a much smaller set of spatially-varying kernels from input, which enlarges the network's receptive field and significantly reduces computational and memory costs. These kernels are then applied to a full-resolution feature map through an efficient slice-and-conv operator with minimum memory overhead. We further build a efficient denoising network using MalleConv, coined as \textbf{MalleNet}. It achieves high quality results without very deep architecture, \eg, it is 8.91$\times$ faster than the best performed denoising algorithms (SwinIR), while maintaining similar performance. We also show that a single \ours added to a standard convolution-based backbones can contribute significantly reduce the computational cost or boost image quality at similar cost. Project page: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2201.00392