Cross-level: A practical strategy for convolutional neural networks based image classification

Abstract

Convolutional neural networks (CNNs) have exhibited great potential in the field of image classification in the past few years. In this paper, we present a novel strategy named cross-level to improve the existing CNNs’ architecture in which different levels of feature representation in a network are merely connected in series. The basic idea of cross- level is to establish a convolutional layer between two nonadjacent levels, aiming to learn more sufficient feature representations. The proposed cross-level strategy can be naturally combined into a CNN without any change on its original architecture, which makes this strategy very practical and convenient. Three popular CNNs for image classification are employed to illustrate its implementation in detail. Experimental results on the dataset adopted by the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) verify the effectiveness of the proposed cross-level strategy on image classification. Furthermore, a new CNN with cross- level architecture is introduced in this paper to demonstrate the value of the proposed strategy in the future CNN design.