Regularized Correlation-Based Integration of Deep Image Descriptors for Remote Sensing Scene Recognition

Abstract

Complex-object recognition in remote sensing imagery is complicated by high intra-class variability, strong inter-class similarity and the heterogeneous statistical properties of deep descriptors extracted from different convolutional models. This paper presents a regularized framework for integrating several image descriptors in a common discriminative space. Label-aware cross-correlation blocks are formed between descriptor matrices, Tikhonov regularization is applied to the block covariance matrix, and projection directions are obtained from a generalized eigenvalue problem. The projected components are concatenated into one integrated descriptor and classified by a support vector machine. Regularization is especially important when the descriptor dimension is comparable to or larger than the training sample size, because covariance blocks may become singular or ill-conditioned. Published benchmark results on NWPU-RESISC45, AID and PatternNet confirm the practical value of correlation-based multi-descriptor fusion for remote sensing scene recognition.