基于YOLOv3-tiny的船舶可见光图像细粒度检测

Fine-grained Detection of Ship Visible Images Based on YOLOv3-tiny

文中针对智能船舶视觉传感器采集到的图像进行目标检测,提出基于YOLOv3-tiny的深度卷积神经网络图像细粒度检测方法.该方法使用实验团队建立的船舶图像数据库,对图片进行人工标注,使用k-means算法对数据集锚框进行聚类,采用数据增强策略的YOLOv3-tiny深度卷积神经网络对船舶图像进行训练与测试.实验结果表明:本文提出的改进YOLOv3-tiny模型在测试集上的平均精度达到了62.85%,实时检测帧率达到了136帧/s,可以辅助船舶驾驶人员识别水面目标.

A fine-grained image detection method based on YOLOv3-tiny was proposed for the images collected by intelligent ship vision sensors.The method used the ship image database established by the experimental team to manually label the images.K-means algorithm was adopted to cluster anchor frames of data sets,and YOLOv3-tiny deep convolution neural network with data enhancement strategy was used to train and test ship images.The experimental results show that the average accuracy of the improved YOLOv3-tiny model in the test set reaches 62.85%,and the real-time detection frame rate reaches 136 f/s,which can assist ship drivers to identify water targets.