Combining Deep and Manifold Learning For Nonlinear Feature Extraction in Texture Images

This paper applies a two-step approach for texture classification by combining Manifold learning with Deep CNN feature extractors. The first step is to use CNN architecture to compute the feature vector of a given image. The second step is to apply Manifold Learning algorithms on the features computed in the first step to making a refined feature vector. Eventually, this final representation is used to train SVM classifier. In the first step, we adopted VGG-19 network trained from scratch in order to extract texture features. In the next step, we used the DIMAL (Deep Isometric Manifold Learning Using Sparse Geodesic Sampling) configuration to train a neural network to reduce the dimensionality of the feature space in a nonlinear manner for generating the refined feature vector of the input image. Our concept is that the combination of a deep-learning framework with manifold learning techniques has the potential to select discriminative texture features from a high dimensional space. Based on this idea, we adopted this combination to perform nonlinear feature extraction in texture images. The resulting learned features were then used to train SVM classifier. The experiments demonstrated that our approach achieved better accuracy in texture classification than existing models if trained from scratch.