结合深度特征迁移与融合的两阶段船牌定位算法

获取运河过往船只的身份信息具有重要意义,快速、准确地定位船牌是实现船只身份自动化识别的首要任务.为提升对小尺度船牌的检测性能,提出一种结合深度特征迁移与融合的两阶段船牌定位算法.首先在船只检测阶段,通过迁移学习构建船只检测模型,获取图片中船只区域的位置信息;然后在船牌定位阶段,提出基于特征融合策略的多尺度船牌定位网络,在上一阶段的基础上对船牌进行定位.在SLPLOC船牌定位数据集上的实验结果表明,相比其他算法,该算法能够有效地减少误差,提升精度值和召回率.

A Two-Stage Ship License Plate Locating Algorithm Based on Deep Feature Transfer and Fusion

It is of great significance to obtain the identity information of the passing ships on the canal.To locate ship license plates quickly and accurately is the primary task to realize automatic identification of ships.In order to improve the performance on detecting small-scale ship license plates,this paper proposes a two-stage algorithm for locating ship license plates based on deep feature transfer and fusion.First,in the stage of detecting ships,a ship detection model is constructed through transfer learning to gain the position information of the ships in the image.Then in the stage of locating ship license plates,a multi-scale locating network with feature fusion strategy is proposed to locate the target plate area based on the previous stage.Experimental results on SLPLOC dataset show that the proposed method can reduce the errors and improve the precision and recall effectively.