MA-SegCloudv1- A novel ground-based cloud image segmentation method based on a multibranch asymmetric convolution module and attention mechanism
With the spirit of reproducible research, this repository contains all the codes required to produce the results in the manuscript:
Please cite the above paper if you intend to use whole/part of the code. This code is only for academic and research purposes.
-
Dataset:The SWINySeg data set is available for download at http://vintage.winklerbros.net/swinyseg.html. All images are normalized to binary images, the size is changed to 320×320, and the training set and test set are divided by voc2pspnet.py.
-
Training: Set the path for train.py to read images and labels, and load training and validation sets.
(a).Read image from file: img = Image.open(r"filepath" + '/' + name + ".jpg")
(b).Read label image from file: label = Image.open(r"filepath" + '/' + name + ".png")
(c).Set hyperparameters such as learning rate, optimizer, loss function, etc.
-
prediction:Load the trained weight file, set the file path to save the prediction results, and run the predict.py.
-
test: Load predictions and labels, run eval.py.