Search and rescue detection algorithm for complex marine environments based on SAN-YOLO

A maritime search and rescue detection algorithm based on SAN-YOLO is proposed to address the issues of lack of high-quality data, weather conditions such as sea fog and rainy days, and poor detection performance of small-sized search and rescue targets in complex maritime environments. Firstly, the SeaDronesSee and self collected dataset are mixed to create a complex marine environment search and rescue detection dataset, which has rich image samples and diverse detection scenarios. Then, a search and rescue detection algorithm for complex marine environments based on SAN-YOLO was proposed, and a backbone network based on SPD-Coord was designed. Through lossless spatial dimensionality reduction and coordinate convolution, the loss of detail information in the feature extraction process was avoided, and the ability to obtain spatial position information was enhanced. Designed a feature fusion network based on C2f-AK attention, which combines C2f and variable kernel convolution to enable the module to have convolution kernels of arbitrary parameters and shapes, enhancing the model’s feature fusion capability. A loss function based on NS Loss was designed, which weighted the sum of NWD Loss and Slide Loss to reduce the sensitivity of the loss function to small target position differences and enhance the weight of complex samples, resulting in better detection performance of the model for small targets. Finally, the algorithm performance was validated on a search and rescue detection dataset in complex marine environments. The experimental results show that the average accuracy of SAN-YOLO is 77.6%, which is 8.9% higher than the benchmark model and better than other common open source algorithms, verifying the effectiveness of the innovative module in this paper. SAN-YOLO has good comprehensive detection performance in search and rescue detection tasks in complex marine environments.