TransYOLO: High-Performance Object Detector for Forward Looking Sonar Images

As a common phenomenon, most modern detectors for forward looking sonar (FLS) images rely heavily on pyramid structure and stream fusion to enhance information and improve performance. But for FLS images with fuzzy boundaries, complex gradients and low information resolution, excavating more potential information representation is necessary. Considering the fuzzy boundaries and complex gradients of FLS images, a transformer feature fusion network (TFFN) based on the transformer stack structure is proposed to promote information fusion. Additionally, we also propose a novel method called ellipse quality evaluation (EQE) to improve the reliability of localization quality estimation, reduce the false detection rate caused by low resolution, and thus improve the detection performance. Finally, we present an anchor-free method based on TFFN and EQE, called TransYOLO. Experiments show that our detector has reached an advanced level in the FLS dataset. Compared with the state-of-the-art detectors, the proposed detector has the best detection performance when it is qualitatively and quantitatively analyzed on the four evaluation metrics in the FLS dataset.