Resolution Learning in Deep Convolutional Networks Using Scale-Space Theory

Resolution in deep convolutional neural networks (CNNs) is typically bounded by the receptive field size through filter sizes, and subsampling layers or strided convolutions on feature maps. The optimal resolution may vary significantly depending dataset. Modern CNNs hard-code their hyper-parameters network architecture which makes tuning such cumbersome. We propose to do away with hard-coded aim learn appropriate from data. use scale-space theory obtain a self-similar parametrization of filters make N-Jet: truncated Taylor series approximate learned combination Gaussian derivative filters. parameter $\sigma $ basis controls both amount detail encodes spatial extent filter. Since continuous parameter, we can optimize it respect loss. proposed N-Jet layer achieves comparable performance when used state-of-the art architectures, while learning correct each automatically. evaluate our classification segmentation, show that especially beneficial dealing inputs at multiple sizes.