A Convolutional Neural Network (CNN) trained on a corpus of images consists of filters tuned to visual features relevant to the task at hand. Variations in the resolution of the images and in the size of the objects and patterns depicted, require the filters to both ignore task-irrelevant scale variations (for recognizing a face, the size of the face is irrelevant) and to respond to task-relevant features at a specific scale (given a scale, the shape and size of the nose are relevant). Previous work focused on developing scale-invariant filters in CNNs. This paper addresses the combined development of scale-invariant and scale-variant filters. We propose a multi-scale CNN method to encourage the development of both types of filters and evaluate it on a challenging image classification task involving task-relevant characteristics at multiple scales. The results show the multi-scale CNN to outperform single-scale CNNs. This leads to the conclusion that encouraging the combined development of scale-invariant and scale-variant filters in CNNs is beneficial to image recognition performance.
View on arXiv