研究目的
To propose a novel change detection framework for SAR images that identifies specific types of changes in semantic level, addressing the limitation of traditional methods that only extract change areas without content identification.
研究成果
The proposed framework effectively detects changes in SAR images at a semantic level, accurately discerning the content of change with high precision and low false alarm rates, outperforming existing methods.
研究不足
Limited to the completeness of the training set, the framework's accuracy usually falls in the edge of the change area.
1：Experimental Design and Method Selection:
The framework combines unsupervised learning (RCAE for feature extraction) and supervised learning (DNN for classification) to detect changes in semantic level. It uses the Bag of Visual Words model for semantic analysis.
2：Sample Selection and Data Sources:
A large number of SAR image slices from harbor areas are used as the training set, including artificial buildings, water, ships, mountain, and other typical content. Test images are from coastal areas with 1m resolution and size 800x800 pixels.
3：List of Experimental Equipment and Materials:
Workstation with Intel I7-8700 CPU and NVIDIA Quadro 4000 GPU, software implemented with Keras and Theano backend, CUDA 8.
4：Experimental Procedures and Operational Workflow:
0.
4. Experimental Procedures and Operational Workflow: SAR images are sliced using a sliding window (size 25x25, step 1 pixel), input to RCAE to get histogram representations, difference vectors are calculated, change areas are segmented using FLICM, and difference vectors are classified by DNN to identify change content.
5：Data Analysis Methods:
Mean square error (MSE) is used to evaluate RCAE performance, classification accuracy is measured via cross validation, and comparison with existing methods is done based on false alarm rate and overall accuracy.
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
