高分辨率遥感影像城市建筑垃圾自动识别的多种标注形式对比研究

Contrastive Study on the Automatic Identification of Urban Construction Waste in HighResolution Remote Sensing Images

数据集的构建是深度学习目标识别中最基础的工作,其在很大程度上决定了目标识别的精度。为研究不同标注形式的数据集对建筑垃圾识别检测效果的影响,分别采用正矩形、旋转矩形和多边形3种标注框的形式标注高分辨率遥感影像中的建筑垃圾,构建了正矩形建筑垃圾数据集、旋转矩形建筑垃圾数据集和多边形建筑垃圾数据集。对比分析实验结果发现:多边形标注框形式的最终识别正确率和识别精准率最高,最适宜建筑垃圾识别模型的构建;利用正矩形和多边形建筑垃圾识别模型均可有效实现对建筑垃圾的自动定位、识别和提取,且多边形建筑垃圾识别模型还可以识别建筑垃圾的轮廓,可以结合实地建筑垃圾高度估算建筑垃圾体积,为建筑垃圾的精准管控提供数据基础和技术支持。

The construction of datasets is the most fundamental task in deep learning target recognition,which largely determines the accuracy of target recognition.To study the impact of different annotation forms of datasets on the recognition and detection performance of construction waste,three types of annotation boxes,namely regular rectangle,rotating rectangle,and polygon,were used to annotate construction waste in high-resolution remote sensing images.The regular rectangle construction waste dataset,rotating rectangle construction waste dataset,and polygonal construction waste dataset were constructed.Comparative analysis of experimental results reveals that the final recognition accuracy and precision of the polygonal annotation box form are the highest,making it the most suitable for constructing a construction waste recognition model.Both rectangular and polygonal construction waste recognition models can effectively achieve automatic positioning,recognition,and extraction of construction waste.The polygonal construction waste recognition model can also recognize the contour of construction waste,estimate the volume of construction waste based on the height of on-site construction waste,and provide data foundation and technical support for precise control of construction waste.