Convolutional Neural Networks (CNNs) require large image corpora to betrained on classification tasks. The variation in image resolutions, sizes ofobjects and patterns depicted, and image scales, hampers CNN training andperformance, because the task-relevant information varies over spatial scales.Previous work attempting to deal with such scale variations focused onencouraging scale-invariant CNN representations. However, scale-invariantrepresentations are incomplete representations of images, because imagescontain scale-variant information as well. This paper addresses the combineddevelopment of scale-invariant and scale-variant representations. We propose amulti- scale CNN method to encourage the recognition of both types of featuresand evaluate it on a challenging image classification task involvingtask-relevant characteristics at multiple scales. The results show that ourmulti-scale CNN outperforms single-scale CNN. This leads to the conclusion thatencouraging the combined development of a scale-invariant and scale-variantrepresentation in CNNs is beneficial to image recognition performance.
展开▼
