Adaptive convolutional network for SAR image classification. Issue 20 (19th September 2019)

• Record Type: Journal Article
• Authors: Xia, Shuang; Yu, Ze; Yu, JinDong
• Is Part Of: Journal of engineering. Volume 2019:Issue 20(2019)
• Page Start: 6868
• Page End: 6872
• Publication Date: 2019-09-19
• Subjects: radar imaging -- neural nets -- learning (artificial intelligence) -- synthetic aperture radar -- image recognition -- image classification -- feature extraction; Adaptive convolutional network -- SAR image classification -- convolutional neural network -- CNN -- highly effective features -- synthetic aperture radar images -- overfitting problem -- sufficient labelled SAR images -- moving target acquisition -- stationary target acquisition -- strong generalisation ability -- different testing conditions -- transfer learning method -- simulated SAR data -- SAR image recognition -- simulated T72 SAR images -- convolution layers -- classification accuracy; Engineering -- Periodicals; Engineering; Electronic journals; Periodicals; 620.005
• DOI: 10.1049/joe.2019.0565
• Languages: English
• ISSNs: 2051-3305

Abstract

Abstract : Convolutional neural network (CNN) has an outstanding performance in image classification via extracting highly effective features automatically. However, when CNN is applied to target recognition of synthetic aperture radar (SAR) images, overfitting problem exists since the lack of sufficient labelled SAR images. Hence, based on the moving and stationary target acquisition and recognition (MSTAR) benchmark dataset, the authors proposed the Adaptive‐Convnet, it has a strong generalisation ability and has a good performance in different testing conditions. Besides, they also studied the transfer learning method to transfer knowledge learned from simulated SAR data to real SAR image recognition, to achieve the purpose of expanding the data set at low cost. When they substitute a part of real T72 images from MSTAR with more simulated T72 SAR images and reduce the dataset bias and enhance transferability in the last three convolution layers, the classification accuracy improved.