研究目的
To propose a novel convolutional neural network architecture for semantic segmentation of bridges with various scales in optical remote sensing images, addressing the challenge of producing dense, pixelwise classification maps for objects with irregular shapes.
研究成果
The proposed network architecture effectively combines semantic information from different layers to produce dense pixel-wise prediction maps for bridge segmentation in remote sensing images. The model demonstrates superior performance and great adaptability, as evidenced by the experimental results.
研究不足
The study is limited by the varying imaging conditions of the remote sensing images, which may affect the model's adaptability and robustness. Additionally, the model's performance is dependent on the quality and diversity of the training dataset.
1：Experimental Design and Method Selection:
The study employs a ResNet as a backbone model to extract semantic features, with a cascaded top-down path added to fuse features at different scales. Joint features are obtained by stacking different layers of feature maps.
2：Sample Selection and Data Sources:
148 high-resolution remote sensing images from Google Earth, with ground sampling distance varying from 0.8m to 1.2m, were collected. Images are sampled over several nations, including Germany, USA, Japan, Korea, Russia, and China.
3：8m to 2m, were collected. Images are sampled over several nations, including Germany, USA, Japan, Korea, Russia, and China.
List of Experimental Equipment and Materials:
3. List of Experimental Equipment and Materials: NVIDIA TITAN X GPU for training and prediction.
4：Experimental Procedures and Operational Workflow:
The model is trained using stochastic gradient descent (SGD) with a learning rate of 10?6, a decay weight of 0.1, and a decay step of 10000. The batch size is set to 5, with 40000 iterations in the dataset.
5：1, and a decay step of The batch size is set to 5, with 40000 iterations in the dataset.
Data Analysis Methods:
5. Data Analysis Methods: Performance is evaluated using pixel accuracy, mean intersection over union (mean IU), and frequency weighted IU.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
