Analysing rotation-invariance of a log-polar transformation in convolutional neural networks

摘要

Applications in computer vision have the challenge of handling objects in images with different orientations and other visual transformations. However, for many tasks, the ideal input feature would be space-invariant regarding geometric transformations, such as angle change, rotation, framing, scale, among others. The simplest way to get invariance to an input class is to train a neural network with augmented data, which does not always capture all changes. In this paper, we propose an architecture of a convolutional neural network that exploits the inherent space-invariance characteristics of the log-polar transformation, which is inspired by the human visual system. We performed experiments on the object classification task and evaluated using several datasets. Our results, employing accuracy metric, show our architecture has the advantage on rotated images, which may be interesting for object detection tasks.