改进CenterNet模型在遥感影像输电杆塔中的应用

Tower Recognition of Satellite Imagery with Improved CenterNet Model

基于卫星影像的输电杆塔的自动识别在电力基础设施规划、建设、维护和灾后评估等方面具有重要意义。但卫星影像因尺度变化差异大、拍摄角度的特殊性、背景复杂多变等问题,给杆塔的自动识别带来了挑战。为此,文章提出了基于改进的CenterNet模型的卫星影像杆塔自动识别算法,一方面在CenterNet骨干网络的输出端增加空间金字塔池化模块,增强模型对多尺度杆塔的适应性;另一方面利用DIoU loss对CenterNet模型的训练过程进行优化。DIoU loss能够直接最小化目标框和预测框之间的距离,使网络收敛速度加快,准确率得到提升。通过对比实验分析,改进后的模型在测试集上的AP指标提升了3%。

Automatic recognition of transmission towers based on satellite images is of great significance in power infrastructure planning,construction,maintenance and disaster assessment.However,due to the differences in the scale of satellite images,the particularity of the shooting angle,and the complex and changeable background,the automatic recognition of tower in satellite images is challenging.Therefore,this paper proposes an automatic recognition algorithm of tower with the improved CenterNet model.On the one hand,the spatial pyramid pool module is added to the output of the CenterNet backbone to enhance the adaptability of the model to multi-scale towers.On the other hand,DIoU loss is used to optimize the training process of CenterNet model.DIoU loss can directly minimize the distance between the target box and the prediction box,so that the network convergence speed is accelerated and the accuracy is improved.Through comparative experiment analysis,the AP index of the improved model is improved by 3%.