| 引用本文: | 郭玲,于海雁,周志权.基于SimAM注意力机制的近岸船舶检测方法[J].哈尔滨工业大学学报,2023,55(5):14.DOI:10.11918/202201069 |
| GUO Ling,YU Haiyan,ZHOU Zhiquan.Nearshore ship detection method based on SimAM attention mechanism[J].Journal of Harbin Institute of Technology,2023,55(5):14.DOI:10.11918/202201069 |
|
| 摘要: |
| 基于视觉图像的船舶目标检测中由于图像背景复杂,无关干扰较多,导致船舶目标检测的难度增大。并且多类别船舶检测数据集现有数量较少且存在样本不均衡的问题使得船舶目标检测性能较低。针对复杂背景干扰检测,本文通过引入SimAM注意力机制对YOLOv3模型进行改进,利用该机制加强船舶目标在提取特征中的权重并抑制背景干扰权重,从而提升模型检测性能;同时,采用强实时数据增强以改善样本尺度分布不均衡的问题,结合迁移学习提升在样本数量受限情况下的船舶检测精度。提取特征的可视化结果显示改进模型对无关背景特征干扰进行了抑制,增强了模型对于船舶特征的提取能力。在SeaShips数据集上,提出的改进模型在不引入额外可学习参数的情况下mAP.5、mAP.75分别达到了96.93%、71.49%,检测速度达到了66 frame/s,在检测精度与运行效率方面保持了均衡。与Saliency-aware CNN、eYOLOv3相比更有效地优化了目标特征,使得mAP.5分别提高了9.53%、9.19%。改进模型在新加坡海事数据集上在船舶类型目标检测的mAP.5达到了81.81%,验证了模型具有较好的泛化能力。 |
| 关键词: 船舶检测 注意力机制 数据增强 改进YOLOv3 迁移学习 |
| DOI：10.11918/202201069 |
| 分类号:TP391.4 |
| 文献标识码:A |
| 基金项目:山东省重点研发计划重大科技创新工程(2020JZZY010705) |
|
| Nearshore ship detection method based on SimAM attention mechanism |
|
GUO Ling,YU Haiyan,ZHOU Zhiquan
|
|
(School of Information Science and Engineering, Harbin Institute of Technology (Weihai), Weihai 264209, Shandong, China)
|
| Abstract: |
| Due to the complex background of ship targets and much irrelevant interference in visual images, it is difficult to conduct ship detection. In addition, there are few datasets for multi-category ship detection and the samples are often unbalanced, which makes the ship target detection performance degraded. Considering the ship detection background interference, an improved YOLOv3 model was proposed by introducing SimAM attention mechanism, which was used to enhance the weight of the ship target in the extracted features and suppress the weight of background interference, thus improving the model detection performance. Meanwhile, strong real-time data augmentation was applied to improve the unbalanced distribution of sample scales, and transfer learning was combined to improve the ship detection accuracy in the condition of a restricted number of samples. The visualization results of extracted features show that the improved model could suppress irrelevant background features, and the abilities of feature extraction and target localization were enhanced. Without introducing additional learnable parameters, the proposed model achieved 96.93% and 71.49% for mAP.5 and mAP.75 on the SeaShips dataset, and detection speed reached 66 frames per second, indicating a good balance between detection accuracy and efficiency. The improved model optimized the target features more effectively compared with the Saliency-aware CNN and eYOLOv3 models, resulting in an improvement of mAP.5 by 9.53% and 9.19%. The mAP.5 for ship type target detection on Singapore Maritime Dataset reached 81.81%, indicating that the proposed model has good generalization performance. |
| Key words: ship detection attention mechanism data augmentation improved YOLOv3 transfer learning |
