Visual attention model based vehicle target detection in synthetic aperture radar images : a novel approach

Gao, Fei and Zhang, Ye and Wang, Jun and Sun, Jinping and Yang, Erfu and Hussain, Amir (2015) Visual attention model based vehicle target detection in synthetic aperture radar images : a novel approach. Cognitive Computation, 7 (4). pp. 434-444. ISSN 1866-9964

[img]
Preview
Text (Gao-etal-CC-2015-Visual-attention-model-based-vehicle-target-detection-in-synthetic-aperture-radar-image)
Gao_etal_CC_2015_Visual_attention_model_based_vehicle_target_detection_in_synthetic_aperture_radar_image.pdf
Accepted Author Manuscript

Download (913kB)| Preview

    Abstract

    The human visual system (HVS) possesses a remarkable ability of real-time complex scene analysis despite the limited neuronal hardware available for such tasks. The HVS successfully overcomes the problem of information bottleneck by selecting potential regions of interest and reducing the amount of data transmitted to high-level visual processing. On the other hand, many man-made systems are also confronted with the same problem yet fail to achieve satisfactory performance. Among these, the synthetic aperture radar-based automatic target recognition (SAR-ATR) system is a typical one, where the traditional detection algorithm employed is termed the constant false alarm rate (CFAR). It is known to exhibit a low probability of detection (PD) and consumes too much time. The visual attention model (VAM) is a computational model, which aims to imitate the HVS in predicting where humans will look. The application of VAM to the SAR-ATR system could thus help solve the problem of effective real-time processing of complex large amounts of data. In this paper, we propose a new vehicle target detection algorithm for SAR images based on the VAM. The algorithm modifies the well-known Itti model according to the requirements of target detection in SAR images. The modified Itti model locates salient regions in SAR images and following top-down processing reduces false alarms by using prior knowledge. Real SAR data are used to demonstrate the validity and effectiveness of the proposed algorithm, which is also benchmarked against the traditional CFAR algorithm. Simulation results show comparatively improved performance in terms of PD, number of false alarms and computing time.