Oceanic Object Detection Based on An Improved SSD Algorithm

In this paper, an optical image object detection method, which combines Single Shot MultiBox Detector (SSD) and attention mechanism, is proposed. Firstly, a typical oceanic object dataset, which contains more than 17,000 images of ships, islands, rail stockades and underwater cables, is constructed. Then, we analyze the advantages of SSD network in multi-scale prediction ability and the reasons for its poor prediction effect on small objects. A detection method, which introduces an attention mechanism after output feature maps of SSD, is proposed. To be specific, this scheme adds CBAM after six output feature maps of SSD. Finally, on the basis of real-time detection, comparative experiments are conducted on oceanic object dataset to select the scheme that can significantly improve the detection accuracy. The results show that the mean average precision (mAP) value under the proposed scheme is 81.90%, and the detection speed is 33.1 frames per second (FPS). Compared to the original SSD, we achieve 3.69% relative improvement on the object detection.