一种基于极端尺度变化的船舶识别方法研究

船舶识别是海上交通监控中非常重要并且具有挑战性的任务,其难度在于复杂场景中对相对较小的船舶进行精确的定位识别。为此提出一种应用于极小船舶目标识别的单级检测算法——YOLO-G算法。由65层卷积层构建特征提取网络;采用多尺度特征融合提取深层语义信息,形成特征金字塔网络执行船舶识别任务;选取先验框机制和调制损失函数来提高识别前/背景的可区分性及模型识别精度。实验使用BOAT数据集和MS-COCO数据集对网络模型进行评估,结果表明,YOLO-G算法性能远高于其他先进的单级检测器,其COCO test-dev@0.5精度值为58.3%。

A BOAT RECOGNITION METHOD BASED ON EXTREME SCALE VARIATION

Boat recognition is an extremely essential and challenging task in marine traffic monitoring.The difficulty of this task lies in the precise positioning and identification of relatively minor boats in complex scenes.Based on the above questions,a single-stage detection algorithm of target recognition of minimal boat—YOLO-G algorithm is proposed.The feature extraction network was constructed by 65 convolution layers;the deep semantic information was extracted by multi-scale feature fusion,and the feature pyramid network was formed to carry out the task of boat recognition;the prior mechanism and modulation loss function were selected to improve the discriminability and model recognition for accuracy of the pre-recognition and background recognition.BOAT dataset and MS-COCO dataset were used to evaluate the network model.The experimental results show that the performance of YOLO-G is much better than that of other advanced single-stage detector,and the precision of COCO test-dev@0.5 is 58.3%.