Learning deep filter banks in parallel for texture recognition

Texture is an important visual clue for various classification and segmentation tasks in the scene understanding challenge. Today, successful deployment of deep learning algorithms for texture recognition leads to tremendous precisions on standard datasets. In this paper, we propose a new learning framework to train deep neural networks in parallel and with variable depth for texture recognition. Our framework learns scales, orientations and resolutions of texture filter banks. Due to the learning of parameters not the filters themselves, computational costs are highly reduced. It is also capable of extracting very deep features through distributed computing architectures. Our experiments on publicly available texture datasets show significant improvements in the recognition performance over other deep local descriptors in recently published benchmarks.