基于改进YOLOv4算法的船舶目标检测方法

Ship Target Detection Algorithm Based on Improved YOLOv4

针对海面环境复杂、船舶目标检测存在检测精度不高和效率低的问题,以及船舶数据集不平衡的现象,提出一种改进YOLOv4算法的船舶目标检测方法。对图像进行预处理,增强船舶图像的有用信息,减少计算量;采用图像增强方法扩充不平衡数据集的小样本数量,提高各类船舶目标检测的准确性;采用改进的K-means++聚类方法重新设计先验锚框,使锚框和目标的边界框更加匹配;采用Softer-NMS对非极大值抑制算法进行优化,对预测框进行后处理,提升模型对密集船舶的检测能力和定位精度。通过开展多组对比试验发现,采用改进的检测算法对10类船舶目标进行识别,精确率P、召回率R和交并比(IOU)等都有很大提高,平均精确率(m AP)值达到96.78%,相比YOLOv4算法提升23.79%;检测速度达到31.2帧/秒,在显著提高检测精度的同时,能缩短检测时间,达到很好的检测效果。

Aiming at the problems of complex sea environment,low detection accuracy and efficiency in ship target detection,and imbalance of ship data set,a ship target detection method based on improved yolov4algorithm is proposed.Firstly,the image is preprocessed to enhance the useful information of the ship image and reduce the amount of calculation;Secondly,the image enhancement method is used to expand the small sample number of unbalanced data set to improve the accuracy of ship target detection;Then,the improved k-means++clustering method is used to redesign the prior anchor frame,so that the anchor frame and the boundary frame of the target are more matched;Finally,softer NMS is used to optimize the non maximum suppression algorithm,and the prediction box is post processed to improve the detection ability and positioning accuracy of the model.Through multiple groups of comparative experiments,the improved detection algorithm has greatly improved the accuracy,recall and IOU of 10 types of ship target recognition,and the map value has reached 96.78%,which is 23.79%higher than yolov4 algorithm;The detection speed is up to 35 frame/s,which not only improves the detection accuracy of yolov4 algorithm,but also shortens the detection time and achieves good detection effect.