YOLO embrace transformer: a two-stage framework for video SAR moving target detection

Video Synthetic Aperture Radar (ViSAR) moving target detection is a challenging task. Generally speaking, the scale of ViSAR moving targets is extremely small, causing the increase of detection difficulties. Existing methods do not perform well in ViSAR small target detection tasks and suffer from loss of information. Moreover, most of the existing methods failed to exploit the intrinsic temporal feature of ViSAR images which is significant for the moving target detection. To address these issues, YOLO Embrace Transformer (YET), a two-stage object detection network is proposed. The first stage network utilizes a model integrated with the Convolutional Block Attention Module (CBAM) which can extract channelwise and spatial-wise information. Meanwhile, fine-grained multi-scale features are merged by inter-channel feature fusion means, enabling the model to detect potential moving targets with a high rate. In the second stage, a classifier is designed to realize false alarm suppression based on temporal information, equipped with a Transformer module. Experiments on both public and private datasets verify the effectiveness of YET. It outperforms mainstream detectors and achieves better precision, indicating its superiority in ViSAR moving target detection task.