Open Vocabulary SAR Object Detection with Denoising and Scaled Box Loss

Synthetic aperture radar (SAR) has important applications in military, geology and other fields due to its high resolution and all-day, all-weather operation. However, the inherent speckle noise of SAR images seriously restricts the effectiveness of detection algorithms. To address this challenge, this paper proposes an open vocabulary SAR object detection model that combines a denoising network and an open vocabulary detection model, named Denoise GroundingDINO (DGD). We add a denoising network AGSDNet before image encoder to remove speckle noise in SAR images and propose a new box loss function called Scaled SCA (SSCA) Loss to improve the model accuracy. Experimental results show that the DGD model has achieved excellent performance on the SSDD, HRSID and SARDet-lOOK (part) datasets, verifying the superior ability of the DGD model in the field of SAR object detection.