【摘要】 针对人工湖中的水面漂浮物检测问题,提出了一种基于改进YOLOv3的水面漂浮物目标检测算法,目标检测包括目标识别与目标定位。首先通过改进的k-means聚类算法获取先验框,以提高定位框与数据集标注框的匹配度,其次在YOLOv3算法框架的3个预测支路中添加类别激活映射(CAM),将原基于边界框的定位方式替换成基于像素点进行定位。实验结果表明:改进的YOLOv3算法提高了识别精度,降低了定位误差。识别精度为97.49%,比YOLOv3算法提高5.14%,平均定位误差为2.60个像素点,比YOLOv3算法减小了1.36。更多还原

【Abstract】 This paper proposes a detection algorithm to detect floating objects based on the improved YOLOv3 algorithm, the concept of detection including identification and localization. An improved k-means clustering algorithm was applied to generate the priori boxes, which improves the matching degree of the detection frame to the target. Besides, the proposed algorithm adds Class Activation Mapping to the three prediction branches in the YOLOv3 frame-work, removes the location boxes, and marks the object location with pixels. The case study has verified that this algorithm improves the accuracy of identification and reduces the deviation in localization. The identification accuracy of this algorithm is 97.49 %, which is 5.14 % higher thanthat of the traditional algorithms. And the average deviation caused in localization is 2.60 pixel point, which decreased by 1.36 compared with the traditional algorithms.更多还原