基于样本重构的船舶小目标检测算法研究

Research on tiny ship detection algorithm based on sample reconstruction

针对遥感影像船舶目标尺度小、场景分布不均匀、目标尺寸相对样本尺寸占比小以及深度学习模型对于小目标泛化性能不好的问题,提出一种样本重构的方法。首先对船舶目标按其最小外接矩形进行裁剪,然后采用多种方式合成标准尺寸样本,通过样本重构,可以提高样本中的目标占比,解决不同场景下目标分布不均匀的问题。实验发现,使用样本重构方法训练的模型,对于小目标的检测能力有所提升,结合在网络中添加小目标检测层,结果显示,模型在测试样本上的平均准确率(average precision, AP)从0.502提升到0.674,验证了该方法的有效性。

Aiming at the problems of small ship target scale,uneven scene distribution,small proportion of target size relative to sample size and poor generalization performance of deep learning model for small targets in remote sensing images,a sample reconstruction method is proposed.Firstly,the ship target is cut according to its smallest circumscribed rectangle,and then a standard size sample is synthesized in various ways.Through sample reconstruction,the proportion of targets in the sample can be increased,and the problem of uneven distribution of targets in different scenarios can be solved.The experiment found that the model trained by the sample reconstruction method has improved the detection ability of small targets.Combined with adding a small target detection layer to the network,the results show that the Average Precision(AP)of the model on the test sample arising from 0.502 to 0.674,verifying the effectiveness of the method.