wr19960001 / siamese-neural-network-for-change-detection Goto Github PK

Implementation of "SIAMESE NETWORK WITH MULTI-LEVEL FEATURES FOR PATCH-BASED CHANGE DETECTION IN SATELLITE IMAGERY" [1] Faiz Ur Rahman, Bhavan Kumar Vasu, Jared Van Cor, John Kerekes, Andreas Savakis, "Siamese Network with Multi-level Features for Patch-based Change Detection in Satellite Imagery", IEEE SigPort, 2018. [Online]. Available:https://sigport.org/documents/siamese-network-multi-level-features-patch-based-change-detection-satellite-imagery. Accessed: Feb. 21, 2019.

We present a patch-based algorithm for detecting structural changes in satellite imagery using a Siamese neural network. The two channels of our Siamese network are based on the VGG16 architecture with shared weights. Changes between the target and reference images are detected with a fully connected decision network that was trained on DIRSIG simulated samples and achieved a high detection rate. Alternatively, a change detection approach based on Euclidean distance between deep convolutional features achieved very good results with minimal supervision.

Dependencies required 1)Tensorflow 2)Keras with tensorflow background 3)Numpy 4)Keras.utils 5)numpy_utils 6)Python 2.7

Data Few sample data in is present in image pairs Unzip the file Names starting with AChip has a corresponding ANeg these are the the pairs for example AChip1,ANeg1 becomes a pair AChip2.ANeg2 becomes a pair

Testing Siamese_predict.py is used for testing open command line and type python Siamese_predict.py It will ask for 1st image chip choose the image pairs as described above Do the same for 2nd image chip Output will be in command line Change or No change Please Cite our work using the bib below.

@inproceedings{rahman2018siamese, title={Siamese Network with Multi-Level Features for Patch-based Change Detection in Satellite Imagery}, author={Rahman, Faiz and Vasu, Bhavan and Van Cor, Jared and Kerekes, John and Savakis, Andreas}, booktitle={2018 IEEE Global Conference on Signal and Information Processing (GlobalSIP)}, pages={958--962}, year={2018}, organization={IEEE} }