一种基于图像显著性的离岸船舶目标检测效率优化方法

An efficiency optimization method of offshore ship target detection based on saliency

目前基于CNN的方法已在高分辨率遥感影像目标检测工作中得到了应用。对像海上船舶这种小型目标的实时检测是该方向研究的难点之一,其主要原因是基于CNN的小型目标检测方法通常伴随较低的检测效率,因此在实时的应用中很难被采用。为此,本文提出了一种以图像显著性为依据的锚点筛选优化方法。该方法充分考虑了海面目标背景的独特性,在对每个像素进行显著性分析的同时,将特征映射中每个锚点对应的接受域进行评分统计。通过显著性机制的运用,使学习和检测过程排除了大量的无效锚点,大幅减少了初始包围窗的生成数量。这种优化过程的主要优势在于它避免了在区域显著性检测时小型船舶目标的流失,而且在训练过程中可以更好地控制正负样本的比例,防止样本不平衡的情况发生。实验证实,本文提出的方法大幅提升了对离岸船舶目标的检测效率,并对基于CNN的两级目标检测方法具有一定的通用性。

At present,CNN based method has been widely applied in object detection of VHR remote sensing images.One of the challenging problems in this field is the detection of small objects,such as offshore ship targets.Since the detection methods of small objects based on CNN are usually accompanied by low detection efficiency,they are difficult to be adopted in real-time applications.To address such problem,a network optimization method based on image saliency score distribution is proposed.By fully considering the peculiarity of the background of offshore ship targets,each pixel is scored by saliency prediction,and then the mean score of each receptive field is obtained corresponding to the anchor in the feature map.Through the application of saliency mechanism,a large number of invalid anchors can be eliminated in the learning and detection process.The main advantage of the proposed method is that,it avoids the loss of small objects in the process of saliency detection,and during the training process,the ratio of positive and negative samples can be better controlled to prevent sample imbalance.Experimental results show that the proposed method significantly improves the efficiency of two-stage frameworks for offshore ship detection.