Detection of Small Targets in Sea Clutter Based on Transformer Recurrence Plots CNN in the Case of Imbalanced Samples

Detecting small targets in sea clutter is challenging due to scarce weak target echoes being overwhelmed by abundance strong sea clutter, leading to imbalanced classification. To address this, we propose a transformer-based data augmentation algorithm that employs a bidirectional prediction mechanism to establish temporal relationships between samples. The augmented data closely align with the actual distribution of target samples, enhancing the target dataset from a 1:10 to a 3:10 ratio with sea clutter. Subsequently, recurrence plots (RPs) of both sea clutter samples and augmented target samples are obtained, and then are input into a convolutional neural network (CNN) for detecting, referred to as Trm-RPs-CNN for short. In order to control false alarm rate of the Trm-RPs-CNN, the output of CNN is replaced by the Softmax probabilities as the detection statistic. This adjustment allows the Trm-RPs-CNN to maintain the high detection probabilities whereas achieve the low false alarm rates by fine-adjusting thresholds. In the experiments, first, we assess the similarity between the augmented target samples and the real target samples by amplitude distribution properties, probability distribution properties, and temporal correlation metrics. Second, seven performance indicators are utilized to evaluate the effectiveness of the Trm-RPs-CNN. Finally, experiments conducted on ten publicly available IPIX radar datasets demonstrate that our proposed Trm-RPs-CNN significantly outperforms traditional approaches in terms of sample balance, classification performance, and false alarm control. In HH polarization, with a false alarm probability of 0.001, the proposed Trm-RPs-CNN achieves an average detection probability of 0.866, demonstrating excellent detection performance.