Gao, Fei and Ma, Fei and Zhang, Yaotian and Wang, Jun and Sun, Jinping and Yang, Erfu and Hussain, Amir (2016) Biologically inspired progressive enhancement target detection from heavy cluttered SAR
F. Gao,Fei Ma,Yaotian Zhang,Jun Wang,Jinping Sun,Erfu Yang,A. Hussain
2018-01-01
Abstract:High-resolution SAR (synthetic aperture radar) can provide a rich information source for target detection and greatly increase the types and number of target characteristics. How to and efficiently extract the target of interest from large amounts of SAR images is the main research issue. Inspired by the biological visual systems, researchers have put forward a variety of biologically-inspired visual models for target detection, such as classical saliency map and HMAX. But these methods only model the retina or visual cortex in the visual system, which limit their ability to extract and integrate targets characteristics, thus their detection accuracy and efficiency can be easily disturbed in complex environment. Based on the analysis of retina and visual cortex in biological visual systems, a progressive enhancement detection method for SAR targets is proposed in this paper. The detection process is divided into RET, PVC, and AVC three stages which simulate the information processing chain of retina, primary and advanced visual cortical, respectively. RET stage is responsible for eliminating the redundant information of input SAR image, enhancing inputs’ features, and transforming them to excitation signals. PVC stage obtains primary features through the competition mechanism between the neurons and the combination of characteristics, and then completes the rough detection. In the AVC stage, the neurons with more receptive field compound more precise advanced features, completing the final fine detection. The experimental results obtained in this study show that the proposed approach has better detection results in comparison with the traditional methods in complex scenes. Keyword: Cortex-Like Mechanisms, synthetic aperture radar (SAR), Hierarchical models, target detection Introduction With the increase of SAR systems and the improvement of their image quality, the targets in SAR images contain richer detail features. The common SAR target detection systems adopt the two-parameter Constant False Alarm Rate (CFAR) algorithm based on Gaussian distribution proposed by Lincoln Laboratory [1]. It takes advantage of differences between the statistical model [2] of background and the targets to achieve targets detection. Whereas, imaging scenes of SAR often contain a large number of artificial buildings, trees, grass$ etc, which form heavy clutters [3] in images and reduce Signal to Clutter Ratio (SCR). The randomness of the clutter makes the estimation of background statistical models more difficult. Moreover, since this method does not make full use of outline, texture, shape and many other features of targets, its detection results are unsatisfied especially under low SCR condition. Biological visual system can search the targets or areas of interests quickly and accurately through properly extracting and processing various kinds of target features, without knowing the model of clutters. Inspired by the unique processing mechanism of the biological vision systems we propose a new idea for the study of SAR target detection. The development of cognitive psychology, neurophysiology, and other related science has unceasingly deepened our understanding for biological vision system [4], which provides the possibility for this attempt. The biological visual system is a complicated multi-level system [5], where low-level areas firstly receive inputs, and then processes these visual information and compound more complex features before transferring them to the higher level. Its main information processing units include retina, primary visual cortex (PVC) and advanced visual cortex (AVC).