Multiscale attention-based detection of tiny targets in aerial beach images

Tiny target detection in marine scenes is of practical importance in marine vision applications such as personnel search and rescue, navigation safety, and marine management. In the past few years, methods based on deep convolutional neural networks (CNN) have performed well for targets of common sizes. However, the accurate detection of tiny targets in marine scene images is affected by three difficulties: perspective multiscale, tiny target pixel ratios, and complex backgrounds. We proposed the feature pyramid network model based on multiscale attention to address the problem of tiny target detection in aerial beach images with large field-of-view, which forms the basis for the tiny target recognition and counting. To improve the ability of the tiny targets' feature extraction, the proposed model focuses on different scales of the images to the target regions based on the multiscale attention enhancement module. To improve the effectiveness of tiny targets' feature fusion, the pyramid structure is guided by the feature fusion module in order to give further semantic information to the low-level feature maps and prevent the tiny targets from being overwhelmed by the information at the high-level. Experimental results show that the proposed model generally outperforms existing models, improves accuracy by 8.56 percent compared to the baseline model, and achieves significant performance gains on the TinyPerson dataset. The code is publicly available via Github.