Detection model optimization method for remote sensing image building, detection method and detection device

The invention discloses a detection model optimization method for a remote sensing image building, a detection method and a detection device. The method comprises the following steps of: optimizing a U-Net network model, specifically, replacing 3 * 3 standard convolution with an asymmetric convolution block in a feature extraction link of a U-Net network, introducing an attention mechanism to a step connection part of the U-Net network to adjust a feature weight; making sample data; inputting the sample data into the transformed U-Net network model for model training; and carrying out precision calculation on a model training result. 本发明公开了一种遥感图像建筑物的检测模型优化方法及检测方法、装置，检测模型优化方法包括如下步骤：优化U-Net网络模型：在U-Net网络的特征提取环节用非对称卷积块代替3×3标准卷积，在U-Net网络的阶跃连接部分引入注意力机制调节特征权重；制作样本数据；将样本数据输入改造后的U-Net网络模型进行模型训练；对模型训练结果进行精确度计算。