A Trustworthy Evaluation Method for Intelligent Target Detection in Marine Backgrounds

The target detection in marine backgrounds faces challenges such as the marine environmental interference and scarce datasets. To evaluate the trustworthiness of intelligent target detection algorithms on marine background datasets, this paper proposes a method based on test data perturbation augmentation and performance boundary evaluation. First, the optical interference, weather perturbations, and temporal perturbations are added to the test dataset, constructing new test sets according to different perturbation intensities. Then, an evaluation index system covering generalization and robustness is established to comprehensively assess the performance and boundaries of target detection algorithms. Experimental results show that under the optical interference, the performance of detection algorithms declines significantly at intensities of 60 % to 80 %; under the weather perturbations, the model performance drops sharply at intensities of 40 % to 50 %; and under the temporal perturbations, the model loses performance after 30% intensity.