SCRIBBLE-SUPERVISED TARGET EXTRACTION METHOD BASED ON INNER STRUCTURE-CONSTRAINT FOR REMOTE SENSING IMAGES
Yitong Li, Chang Liu, Jie Ma*
2023 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 16-21 July 2023
Paper
''' sh preprocess_dataset.sh '''
Note: After running the above code, you will get a directory named "examples" where input images are stored in 'examples/images/{image_type}/' and then loaded according to 'examples/labels/{image_type}/train.csv' and 'examples/labels/{image_type}/test.csv' for trainning and testing respectively. Besides, scribble labels in which target region, background and unknown area are correspondingly annotated by pixels with values of 1, 0 and 250 are generated and stored in 'examples/labels/{image_type}/'. To apply other datasets, you can modify "preprocess_dataset.sh" and "preprocess.py". In testing, fully-annotated labels in 'examples/GT/{image_type}/' are used for validation, which need to be manually created. The airplane data used in the paper are airport satellite images from Google, and the airplanes and background are annotated sparsely in our scribbles.
(1) Modify the paths in "TRAIN" of model.json ("config/model.json").
(2) ''' python Train.py --config_path config/model.json '''
(1) Modify the paths in "TEST" of model.json ("config/model.json").
(2) ''' python Inference.py --config_path config/model.json '''
(1) Modify the paths in model.json ("config/model.json").
(2) ''' sh run.sh '''
In the paper, the hyper-parameters "num" and "degree" in "TRAIN" are set to 0.3 and 0.7, which has been decided after enough comparative experiments. Prediction results in testing with different hyper-parameter settings are partly shown below, proving the effectiveness of our setting in the paper.
Input image
"num":0.1, "degree":0.7
"num":0.3, "degree":0.7
"num":0.5, "degree":0.7
"num":0.7, "degree":0.7
"num":0.9, "degree":0.7
"num":0.3, "degree":0.1
"num":0.3, "degree":0.3
"num":0.3, "degree":0.5
"num":0.3, "degree":0.9
