Keras(TF backend) implementation of unet for RGB pictures.

• OpenCV
• Python 2.7
• Tensorflow-gpu
• Keras

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

• extract downloaded data to corresponding directories
• run python data.py to generate 3 .npy files or you can download them to npydata
• run python unet.py to train, you can change hyperparameters on your own situation
• run python 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.

• run python data_camvid.py
• run python unet_camvid.py
• run 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.

https://github.com/zhixuhao/unet