建筑垃圾堆放点样本集构建与优化研究

为更好地解决手工制作的建筑垃圾堆放点样本集效率低、数据量少，难以支撑基于深度学习的遥感图像目标检测算法训练需求的问题，采用基于像素的遥感分类方法构建建筑垃圾堆放点样本集，在此基础上结合直方图均衡化，CS-LBP算子约束以及迁移学习的方法对Wasserstein生成对抗模型（WGAN）进行优化，实现了样本集扩充。研究结果表明:相对于纯手工制作的样本集，基于像素的遥感分类方法可以显著提升样本集制作的效率；同时，经过WGAN优化后，生成样本模拟了原始数据的颜色与纹理特征分布规律，增加了原始数据的多样性，满足了扩充样本集的需求。

CONSTRUCTION AND OPTIMIZATION OF SAMPLE SET OF CONSTRUCTION WASTE DUMP

The unregulated stacking of urban construction waste endangered the environment and the safety of citizens. In order to identify the stacking points of construction waste, aiming at the problem that there was no stacking point sample set for construction waste, this paper used the pixel based remote sensing classification method to build a sample set. On this basis, the adaptive histogram averaging, CS-LBP operator constraints and migration learning were combined to optimize the Wasserstein generative adversarial networks(WGAN) and generate samples to expand the sample set. The experimental results showed that the pixel based remote sensing classification method improved the efficiency of the artificial sample set, and the WGAN optimized method could effectively inherit the color and texture features of the original data to meet the needs of expanding the sample set.