Scale Equivariant CNNs with Scale Steerable Filters
Paper ID : 1120-MVIP2020
hanieh Naderi *, leili goli, Shohreh Kasaei
Convolution Neural Networks (CNNs) despite being one of the most successful image classification networks, are not robust to most geometric transformations (rotation, isotropic scaling, etc.) because of their structural constraints. Recently, scale steerable filters have been proposed to allow scale invariance in CNNs. Although these filters enhance performance of the network in scaled image classification task, they can not maintain the scale information across the network. In this paper, this problem is addressed. First, a CNN is built with the usage of scale steerable filters. Then the core idea of this paper aiming to acquire a scale equivariat network, is implemented by adding a feature map to each layer so that the scale-related features are retained across the network. At last by defining the cost function as cross entropy, this solution is evaluated and the model parameters are updated. This network is evaluated on MNIST-scale and FMNIST-scale datasets. The results show that the proposed method on the FMNIST-scale dataset has about 2% better performance than other comparable methods for scale equivariance and scale invariance. One of the limitations of this study is that only two datasets have been used to test network robustness. Furthermore only scale transformation is aimed and the network remains vulnerable to many other geometric transformations. It is assumed that in FMNIST and MNIST datasets all images are approximately in one similar base scale originally.
Equivariance, Invariance, Convolutional Neural Networks, Scale , Steerable Filter close
Status : Paper Accepted (Oral Presentation)