Abstract: Aiming at the low detection rate caused by complex shipboard scenes and mutual occlusion of targets, an enhanced YOLOv3 algorithm suitable for shipboard target detection is proposed based on the YOLOv3 algorithm. The integrated data augmentation strategy is added to the input network to carry out gamut transformation, clipping, shielding and other operations on the image, and then a variety of image selection, transformation and combination methods are designed to enrich the sample information. According to the characteristics of the target size on ship surface, K-means algorithm is used to redesign the prior box matching the detection target and allocate it to the corresponding prediction scale. In the output network, the parameter setting of Gauss soft threshold function of Soft-NMS algorithm is improved by linear function to meet the need of suppression under different densities. Through enhanced detection algorithm in the target data set on experimental comparison, the results show that five types of ship surface target are identified, and precision rate and recall rate are increased by 1.4% and 10.3% respectively, and the average accuracy value（mAP） reaches 95.24%, and then detection speed is 21.5 frame/s, which effectively solves the complicated situations of ship surface much target detection problem.

Key words: YOLOv3 algorithm, K-means, non-maximum suppression（NMS）, dense target detection, data augmentation

摘要： 针对舰面场景复杂、目标相互遮挡导致检测率较低等问题，在YOLOv3算法基础上提出了适用于舰面目标检测的增强YOLOv3算法。在输入网络中加入融合的数据增强策略对图像进行色域变换、裁剪、遮挡等操作，设计了多种类图片选取、变换及组合方式来丰富样本信息；针对舰面目标尺寸的特点，利用K-means算法重新设计与检测目标相匹配的先验锚框并分配至对应的预测尺度，以加速模型收敛；在输出网络中通过线性函数对Soft-NMS算法的高斯软阈值函数参数设定进行了改进，以适应不同密集度下的抑制需要，提高网络检测能力。通过将增强的目标检测算法在目标数据集上进行实验对比，其结果显示，在5类舰面目标识别的精确率和召回率分别提高了1.4%和10.3%，平均准确率值（mAP）达到了95.24%，检测速度达到21.5?frame/s，有效解决了复杂场景下的舰面多目标检测问题。

关键词: YOLOv3算法, K-means, 非极大抑制, 密集目标检测, 数据增强