Abstract:

In the processing of vehicle-target detection in synthetic aperture radar (SAR) images,the contours of vehicles not only provide their position but also represent their condition,which is a key to SAR image understanding.But the multiplicative speckle noise in SAR images interferes with the border positioning of vehicles,resulting in difficulties for vehicle-target detection.To solve this problem,the present paper proposes an attention-mechanism-based neural network for pixel level vehicle detection,which consists of a target filtering module,a target locating module and a contour refining module.The target filtering module contains a lightweight feature extraction network with a channel-attention and self-attention mechanism to enhance feature expression.This module can decrease the effect of the speckle on features to select images containing the target quickly and precisely,and provide the output stable location heat map for the next module.The target locating module uses the foreground-background cross-attention mechanism to refine the coarse-scale features in accordance with the location heat map and refine the target location.Furthermore,the module adopts the fine-scale information to improve the details of the target contour.The contour refining module eliminates the contour uncertain points caused by upsampling and speckle noise to obtain accurate contour pixel confidence.For testing this network,a target image dataset and a large-scene image dataset are built with the pixel-level vehicle annotation of the dataset labeled by ourselves.The result of testing indicates that the network has a good pixel-level detection performance and can detect vehicle targets in large SAR images rapidly and accurately.

Key words: vehicle detection, deep learning, attention mechanism, synthetic aperture radar(SAR), pixel-level target detection