Parametric Learning of Texture Filters by Stacked Fisher Autoencoders

Deep learning has recently contributed significantly to large-scale recognition of several modalities like image, video and speech. Stacked autoencoders are a family of powerful convolutional neural nets to build scalable generative models for automatic feature learning. In this paper, we propose a network of novel overcomplete autoencoders called Fisher autoencoders. In contrast to convolutional autoencoders which learn some latent representations, we train a set of projections for the model variables using banks of filters. The Fisher autoencoders are independently computed in stacks of variable depth based on the complexity of patterns under study and the ability of each individual filter to extract deep features. We select texture understanding as one of the most difficult tasks in pattern recognition and conduct our experiments in a standard platform to assure fair comparisons with other methods. Our results show considerable improvements over the most recent benchmarks on several texture datasets for our Fisher autoencoders evaluated against improved Fisher vectors on Dense SIFT (DSIFT) and DeCAF-VGG deep local descriptors.