Learning of Separable Filters by Stacked Fisher Convolutional Autoencoders

Learning of convolutional filters in deep neural networks proves high efficiency to provide sparse representations for the purpose of image recognition. The computational cost of these networks can be alleviated by focusing on separable filters to reduce the number of learning parameters. Autoencoders are a family of powerful deep networks to build scalable generative models for automatic feature learning. Inspired by their stacked hierarchy, we introduce Fisher convolutional autoencoders to learn separable filters in a distributed architecture. These novel overcomplete autoencoders employ discriminant analysis to impose the highest possible distinction among texture classes whilst holds the minimum separation within each individual class. A distributed network of stacked Fisher autoencoders learns banks of separable filters in parallel and makes an ensemble of deep convolutional features with higher separability for a better classification. This network automatically adjusts depth of each stack with respect to the capability of its correspondent separable filter on extracting higher order convolutional features for the dataset under study. We conduct our experiments on several publicly available datasets varying in number of classes and quality of samples by using a standard implementation. Our results confirm the supremacy of our method on improving the precision of texture understanding in comparison with the recently published benchmarks.