YOLF-ShipPnet: Improved RetinaNet with Pyramid Vision Transformer

Abstract

Abstract In the field of ship detection, the intricate nature of ship images arises from a multitude of factors, including variations in ship orientation, color contrasts, and diverse shapes. These factors collectively contribute to the challenge of achieving high detection precision. Thus, it is necessary to investigate the application of advanced networks for ship image detection. In this paper, we have put forward an improved network called YOLF-ShipPnet, which utilizes a popular pyramid vision transformer with increased depth as the backbone for the RetinaNet network. To increase the model’s generalization ability, You Only Look Once eXtreme’s (YOLOX’s) hue, saturation, and value (HSV) random augmentation technique is employed to simulate light and color effects on ship images during the construction of the network. Ablation experiments were conducted on the model with two popular datasets: High-Resolution Ship Collections 2016 (HRSC2016) and SAR Ship Detection Dataset (SSDD). The YOLF-ShipPnet network has been verified to improve detection precision and generalization ability in ship detection by $$5.22\%$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mn>5.22</mml:mn> <mml:mo>%</mml:mo> </mml:mrow> </mml:math> and $$5.46\%$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mn>5.46</mml:mn> <mml:mo>%</mml:mo> </mml:mrow> </mml:math> , respectively, compared to RetinaNet baseline, exhibiting strong robustness and high effectiveness. The proposed network is applicable to the field of fine-grained ship detection and achieves an accuracy improvement of $$10.03\%$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mn>10.03</mml:mn> <mml:mo>%</mml:mo> </mml:mrow> </mml:math> compared to the baseline network.