Deep learning methods have recently made a significant breakthrough in many classification problems. Deep convolutional networks realize the possibility of building large-scale data representations with multiple processing layers. A scattering transform is an effective and relatively recent deep convolutional network that builds large-scale informative representations by cascading linear and nonlinear operators in multiple layers. However, experimental results show that scattering coefficients have a significantly low energy at higher layers and subsequently have little impact on classification results. In this paper, we propose a multi-wavelet fusion-based architecture to improve the error of classification by using scattering transforms. The method effectively increases the energy of scattering coefficients at higher levels via a further step of applying a linear operator followed by a nonlinear operator. Our experimental results showed that the proposed method achieves better classification results compared to the original scattering transform, and thus can be used as an effective method for texture classification problems.

Date of Conference: 13-15 February 2017

Publisher: IEEE

Conference Location: Churchill, VIC, Australia