unet for rgb images semantic segmentation
Keras(TF backend) implementation of unet for RGB pictures.
You can download all data here:https://pan.baidu.com/s/1WqmR-9jodyyUnLRCK6pcVw password: 533h
CamVid Data:https://github.com/preddy5/segnet/tree/master/CamVid
python data.py
to generate 3 .npy files or you can download them to npydatapython unet.py
to train, you can change hyperparameters on your own situationpython test2mask2pic.py
to get test results(pictures)After training about 30 epochs, loss goes to about 0.005.The results seems OK! But the edges look a bit rough. I think that is Unet's own limitation.
python data_camvid.py
python unet_camvid.py
python predict_camvid.py
After 50 epochs training, loss goes to about 4e-04 and acc is 0.9644.
Unet is mostly used in medical areas. I used this model for semantic segmentation of satellite remote sensing images in real work and the result is not bad. I think one of the reason is that it's a coarse-grained task like medical image analysis. Of course, enough labeled images are necessary.