内河通航环境日益复杂，利用航道视频监控系统对船舶图像进行目标识别，对船舶安全监测具有重要意义.内河不同类型船舶在尺度上具有显著差异，且行驶在航道不同位置的船舶相对于视频监控端存在视距差异，使得船舶目标在成像中呈现出多尺度现象，特别是小尺度船舶图像目标呈现的特征信息较少，导致现有YOLOv3算法对小尺度船舶识别精度较低.针对YOLOv3算法中IoU边界框回归损失函数对小尺度船舶预测框容易产生误识及漏识的问题，利用目标预测框与真实框的最小闭包区域和关键点的归一化距离，提出一种新的损失函数MIoU，该损失函数能显著提升多尺度船舶目标预测框的回归速度及精度.实验表明：提出的YOLOv3-MIoU算法对六类船舶的识别精度均超过了97%，mAP值达到了98.44%.与采用其他损失函数的方法相比，YOLOv3-MIoU在不同尺度及不同类型船舶识别中均具有较高的识别准确率，特别在渔船等小尺度船舶识别准确率比其他方法提升了3%以上，可以达到内河航运安全监控的应用需求.

Aiming at the problem that the regression loss function of IoU boundary box in YOLOv3 algorithm was prone to misidentification and missing identification of smallscale ship prediction box, a new loss function MIoU was proposed by using the normalized distance between the minimum closure area and key points of the target prediction box and the real box, which can significantly improve the regression speed and accuracy of multiscale ship target prediction box. Experiments show that the proposed YOLOv3MIoU algorithm has a recognition accuracy of more than 97% for the six types of ships, and the mAP value reaches 98.44%. Compared with other loss function methods, YOLOv3MIoU has higher recognition accuracy in different scales and types of ship image targets, especially for smallscale ships such as fishing boats, the recognition accuracy is improved by more than 3% higher than other methods, which can meet the application needs of inland shipping safety monitoring.