A sea–sky–line detection method for long wave infrared image based on improved Swin Transformer

Abstract

Long wave infrared (LWIR) imaging technology is booming due to its all-weather capability. Sea–sky–line (SSL) detection based on LWIR images is a promising research in marine environment perception. However, LWIR images have been suffering the lack of rich features, challenge arises from the accurate SSL detection in complex sea–sky background. In this paper, we propose a novel SSL detection method for LWIR images, which consists of three algorithms. First, a three-channel reconstruction algorithm for local images is proposed to increase the amount of SSL features. Second, an improved Swin Transformer network is presented for local image SSL identification, which improves the identification speed while ensuring accuracy. Third, a local SSL extraction algorithm is designed and applied to global SSL detection. Experimental results demonstrate that the proposed SSL detection method is more robust to complex background environments than the existing methods. As high as 98.9% average accuracy of SSL detection in LWIR images can be achieved, which outperforms all comparison methods. The extracted SSL is visually closer to the nature SSL, where the radian effect caused by camera distortion can be well fitted. Moreover, ablation studies are also conducted to validate the effectiveness of the proposed three algorithms in our method.