Adaptive-scale convolutional neural networks for texture image analysis

Abstract

This paper proposes an effective adaptive-scale convolutional neural networks (A-SCNN) for texture image analysis. We combine the multi-scale texture image analysis with the efficient feature space of a convolutional neural network to extract characteristic texture features. These latter encode regions of adaptive sizes centered on each pixel according to different optimal scales reflecting the local structure pattern content. To fix the scale-space values accurately, the Hessian-Laplacian operator is used. Experimental results demonstrate a good performance of the proposed A-SCNN in texture classification. Particularly, the CNN based on the adaptive scale shows promising for irregular texture pattern classification, and the selective sizes of both feature maps and receptive fields can further improve the performance of the classical CNN texture discrimination ability.