Deep invariant feature learning for remote sensing scene classification

Wang, Shidong (2021) Deep invariant feature learning for remote sensing scene classification. Doctoral thesis, University of East Anglia.

[thumbnail of 2020WangSPhD.pdf]
Download (36MB) | Preview


Image classification, as the core task in the computer vision field, has proceeded at a break­neck pace. It largely attributes to the recent growth of deep learning techniques which have blown the conventional statistical methods on a plethora of benchmarks and even can outperform humans in specific image classification tasks. Despite deep learning exceeding alternative techniques, they have many apparent disadvantages that prevent them from being deployed for the general-purpose. Specifically, deep learning always requires a considerable amount of well-annotated data to circumvent the problems of over-fitting and the lacking of prior knowledge. However, manually labelled data is expensive to acquire and is impossible to incorporate the variations as much as the real world. Consequently, deep learning models usually fail when they confront with the underrepresented variations in the training data. This is the main reason why the deep learning model is barely satisfactory in the challeng­ing image recognition task that contains nuisance variations such as, Remote Sensing Scene Classification (RSSC).

The classification of remote sensing scene image is a procedure of assigning the seman­tic meaning labels for the given satellite images that contain the complicated variations, such as texture and appearances. The algorithms for effectively understanding and recognising remote sensing scene images have the potential to be employed in a broad range of applications, such as urban planning, Land Use and Land Cover (LULC) determination, natural hazards detection, vegetation mapping, environmental monitoring. This inspires us to de­sign the frameworks that can automatically predict the precise label for satellite images. In our research project, we mine and define the challenges in RSSC community compared with general scene image recognition tasks. Specifically, we summarise the problems into the following perspectives. 1) Visual-semantic ambiguity: the discrepancy between visual features and semantic concepts; 2) Variations: the intra-class diversity and inter-class similarity; 3) Clutter background; 4) The small size of the training set; 5) Unsatisfactory classification accuracy in large-scale datasets.

To address the aforementioned challenges, we explore a way to dynamically expand the capabilities of incorporating the prior knowledge by transforming the input data so that we can learn the globally invariant second-order features from the transformed data for improving the performance of RSSC tasks. First, we devise a recurrent transformer network (RTN) to progressively discover the discriminative regions of input images and learn the corresponding second-order features. The model is optimised using pairwise ranking loss to achieve localising discriminative parts and learning the corresponding features in a mutu­ally reinforced way. Second, we observed that existing remote sensing image datasets lack the provision of ontological structures. Therefore, a multi-granularity canonical appearance pooling (MG-CAP) model is proposed to automatically seek the implied hierarchical structures of datasets and produced covariance features contained the multi-grained information. Third, we explore a way to improve the discriminative power of the second-order features. To accomplish this target, we present a covariance feature embedding (CFE) model to im­prove the distinctive power of covariance pooling by using suitable matrix normalisation methods and a low-norm cosine similarity loss to accurately metric the distances of high­dimensional features. Finally, we improved the performance of RSSC while using fewer model parameters. An invariant deep compressible covariance pooling (IDCCP) model is presented to boost the classification accuracy for RSSC tasks. Meanwhile, we proofed the generalisability of our IDCCP model using group theory and manifold optimisation techniques. All of the proposed frameworks allow being optimised in an end-to-end manner and are well-supported by GPU acceleration. We conduct extensive experiments on the well-known remote sensing scene image datasets to demonstrate the great promotions of our proposed methods in comparison with state-of-the-art approaches.

Item Type: Thesis (Doctoral)
Faculty \ School: Faculty of Science > School of Computing Sciences
Depositing User: Chris White
Date Deposited: 08 Jun 2021 13:19
Last Modified: 08 Jun 2021 13:19

Actions (login required)

View Item View Item