High-efficiency scene classification based on deep compressed-domain feature

• Main Authors: Cheng Li, Baojun Zhao, Boya Zhao, Wenzheng Wang, Chenhui Duan
• Format: Article
• Language: English
• Published: Wiley 2019-08-01
• Series: The Journal of Engineering
• Subjects: remote sensing; image coding; data compression; geophysical image processing; image classification; feature extraction; image representation; convolutional neural nets; computing resources; in-orbit processing capacity; deep feature representation framework; body information; raw JPEG2000 codestream; compressed-domain convolutional neural network; high-level representation; effective classification; high classification accuracy; high-efficiency scene classification; compressed-domain feature; remote sensing image scene classification; RSI; satellite application; header information; UAV time-sensitive application
• Online Access: https://digital-library.theiet.org/content/journals/10.1049/joe.2019.0266

Remote sensing image (RSI) scene classification has become a more and more fundamental issue in satellite and UAV time-sensitive applications. However, as the volume and velocity of RSIs are undergoing an explosive growth, traditional effective technologies claim a huge amount of computing resources, on a scale that already goes beyond in-orbit processing capacity. Here, we attempt to design a deep feature representation framework based on onboard compressed data to solve the aforementioned problem. Firstly, we extract header and body information in raw JPEG2000 codestream for representing geometrical and geospatial property of RSIs. Moreover, a novel compressed-domain convolutional neural network (CNN) is proposed to obtain high-level representation for effective classification. Extensive experiments demonstrate that, in comparison with the existing relevant state-of-the-art approaches, the proposed method achieves high classification accuracy with faster computation and lower consumption.