基于弱语义注意力的遥感图像可解释目标检测

Weakly Semantic Based Attention Network for Interpretable Object Detection in Remote Sensing Imagery

近些年来随着遥感技术的快速发展,遥感图像目标检测成为了当前的研究热点.针对遥感图像背景复杂以及现有目标检测模型缺乏可解释性等问题,本文提出了一种基于弱语义注意力的遥感图像可解释目标检测方法.具体地,首先通过多层级特征金字塔来解决遥感图像中目标尺度变化范围大的问题.其次,利用检测框的角度回归来解决遥感图像目标定向的问题.然后,基于弱语义分割网络产生强化目标特征的注意力权重值,抑制背景噪声.最终用网络剖析的分析方法,获取模型中卷积核对应的可解释性语义概念.实验结果表明,本文提出的算法在遥感图像目标检测的准确性以及对背景噪声抑制上有较好的表现,并且通过可解释性算法在一定程度上使本文提出的模型易于理解.

In recent years,the object detection in remote sensing imagery has been a hot research spot with the development of remote sensing technique.To deal with the complex background in the imagery and the detection model’s interpretability,we propose a weakly semantic based attention network for interpretable object detection model in remote sensing imagery.Firstly,a feature pyramid network is devised for the variation of object scales.Next,an angle is added to the regression to better locate the object.Thirdly,we add a weakly semantic segmentation network to enhance feature and filter the noisy information in the background.Finally,the model is dissected by the proposed method to get the interpretable semantic concepts of convolutional kernel.Experiments validated that the model has a good performance in the aspect of suppressing the background noise and make our model easy to comprehend.