Receptive Field Convolutional Neural Networks and Applications in Image Classification
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Convolutional Neural Networks (CNN) have reached an impressive performance in object detection and classification tasks. However, such success requires massive amounts of high-quality labeled data and this is impractical in applications such as medical and hyperspectral imaging, where data annotation is labor-intensive and often requires domain experts. This dissertation presents a novel strategy to reduce the need for massive labeled data based on Receptive Field Convolutional Neural Network (RFCNN), a new class of CNNs, where convolutional filters are selected as linear combinations from a predefined dictionary and only the coefficients of this combination need to be learned. Our main contributions include the introduction of a sparsity constraint in combination with a recently introduced family of redundant framelets dictionary to reduce the number of parameters of the network while improving stability and generalization. To illustrate our approach, we consider problem of image classification using three challenging datasets: the UH 2013 hyperspectral dataset from the IEEE GRSS Data Fusion Contest, the Quick, Draw! Doodle Recognition Challenge dataset, and the Imagenet Large Scale Visual Recognition Challenge 2012 dataset. In addition, we propose a new deep learning strategy for a specific Hyperspectral Image classification task, namely the UH 2018 hyperspectral dataset from ”2018 IEEE GRSS Data Fusion Challenge”, where we integrate different deep learning strategies to efficiently learn joint spatial-spectral features over multiple scales.