Differencing Neural Network for Change Detection in Synthetic Aperture Radar Images

This paper presents a completely unsupervised change detection approach for synthetic aperture radar (SAR) images based on stacked autoencoders (SAE). The proposed method innovatively implements the change detection task by establishing a differencing neu

  • PDF / 200,976 Bytes
  • 7 Pages / 439.37 x 666.142 pts Page_size
  • 34 Downloads / 257 Views

DOWNLOAD

REPORT


School of Electronics and Information, Northwest Polytechnical University, 127 West Youyi Road, Xi’an 710072, Shaanxi, People’s Republic of China [email protected], [email protected] 2 Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of China, Xidian University, 2 South TaiBai Road, Xi’an 710071, Shaanxi, China [email protected]

Abstract. This paper presents a completely unsupervised change detection approach for synthetic aperture radar (SAR) images based on stacked autoencoders (SAE). The proposed method innovatively implements the change detection task by establishing a differencing neural network with a novel cost function. Firstly, two SAR images are used to pre-train two stacked autoencoders, then these two stacked autoencoders are unrolled to initialize the parameters of differencing neural network. Next, a novel cost function, including the difference between bi-temporal features and an initial difference image, is designed to fine tune the networks for highlighting the changes. Finally, we can obtain the detection results by measuring the Euclidean distance between the outputs of the two neural networks. The experiments on real multitemporal SAR datasets prove the outstanding performance of the proposed method. Keywords: Differencing neural network · Image change detection · Stacked autoencoder · Cost function · Synthetic aperture radar (SAR)

1

Introduction

Image change detection is to recognize the changes between two images which are taken over the same scene but at different times. Because of the independence of time and weather condition, synthetic aperture radar (SAR) image has received a lot of attention in recent years and is the main experimental object of the change detection algorithm. Nevertheless, due to the presence of the speckle noise, it is more difficult to achieve the change detection for SAR image [1]. In the last few decades, the common way to handle change detection for SAR images is post-comparison analysis according to the existed literature [2]. It is called difference image (DI) analysis [3]. However, the quality of DI will greatly affect result of change detection. In recent years, with the rise of neural c Springer Nature Singapore Pte Ltd. 2016  M. Gong et al. (Eds.): BIC-TA 2016, Part I, CCIS 681, pp. 431–437, 2016. DOI: 10.1007/978-981-10-3611-8 38

432

F. Chen et al.

networks, researchers have tried to use the neural networks to solve change detection problems, such as a Hopfield-type neural network proposed for multispectral images [4]. This paper presents an unsupervised change detection approach for SAR images based on stacked autoencoder (SAE). To implement the aim of change detection, a novel cost function is designed to adjust the parameters of two coupled neural networks. Finally the trained deep neural networks are used for the classification of changed pixels by two original images directly. The whole framework is called differencing neural network by ourselves. The rest of this paper is organized as follows. Section 2 will desc