Robust Neural Network Training for Marine Radar Target Detection Using Noisy AIS Labels

For the neural network (NN) detector in marine radar, its superior performance relies heavily on correctly annotated large datasets bringing expensive cost. The automatic identification system (AIS) can be used to automatically construct datasets by obtaining the location of other ships. However, the constructed dataset inevitably contains a large number of noisy labels, which can seriously affect the performance of the NN. Therefore, we propose a robust NN training method for marine radar ship target detection using noisy AIS labels. First, we propose the radar-assisted AIS label construction strategy to automatically construct datasets with noisy AIS labels. Second, we analyze and find that the differentiable Neyman Pearson (DNP) loss is robust to noisy AIS labels due to its relaxing effect on incorrect labels and reinforcing effect on correct labels, and propose a dynamic adjustment strategy to make the DNP loss robust to noisy AIS labels with an unknown number of error labels. Third, to effectively train the NN detector with noisy AIS labels and the DNP loss, we propose the logit-peak-distillation network training strategy, where we use the NN logit value processing strategy to quickly switch to lower probability of false alarm (PFA) by avoiding the information loss caused by the sigmoid activation function, use the peak detection strategy to accelerate the NN training by efficiently computing the PFA, and use the ensemble knowledge distillation strategy to reduce the effect of error labels by cleaning the noisy AIS labels into cleaner soft labels. In the experiment, compared with the benchmark methods, our proposed method has the better detection performance on the measured data of marine radar.