U-Net is pixel level classification, and the output is the category of each pixel, and different categories of pixels will display different colors. U-Net is often used in biomedical images, and the image data in this task is often small. Therefore, Ciresan et al. trained a convolutional neural network to predict the class tag of each pixel using the sliding window to provide the surrounding area (patch) of the pixel as input.

I didn't find any pretrain weights. It's also different to other network like FCN, SegNet, doesn't have familiar backbone. U-Net Paper said it used isbi dataset, but I didn't complete reproduce U-Net which has more suitable for medical image padding manipulation. You can think of it as a simple U-Net. So I just random initialize weights to train in VOC dataset.

I still implement ISBIdataset.py. It has elastic_transform function which is a data augmentation for cell membranes described in this paper.

1. Pull this repository.

git clone https://github.com/verages/UNet-keras.git

2. You need to install some dependency package.

cd U-Net-keras
pip installl -r requirements.txt

3. Download the VOC dataset(VOC SegmetationClassAug if you need).
4. Getting U-Net weights.

wget 

5. Run predict.py, you'll see the result of U-Net.

python predict.py

Input image:

Output image（resize to 320 x 320）:

1. You should rewrite your data pipeline, Dateset where in dataset.py is the base class, such as VOCdataset.py.

class VOCDataset(Dataset):
    def __init__(self, annotation_path, batch_size=4, target_size=(320, 320), num_classes=21, aug=False):
        super().__init__(target_size, num_classes)
        self.batch_size = batch_size
        self.target_size = target_size
        self.num_classes = num_classes
        self.annotation_path = annotation_path
        self.aug = aug
        self.read_voc_annotation()
        self.set_image_info()

2. Start training.

python train.py

3. Running evaluate.py to get mean iou and pixel accuracy.

python evaluate.py
--------------------------------------------------------------------------------
Total MIOU: 0.4472
Object MIOU: 0.4243
pixel acc: 0.8939
IOU:  [0.90583348 0.61631588 0.46176812 0.32607745 0.20159544 0.33146206
 0.73923587 0.61883991 0.61589186 0.22619494 0.26843258 0.28243826
 0.47725354 0.40103127 0.50151519 0.6797911  0.20051836 0.33062457
 0.28617971 0.49376667 0.42613842]

4. If you can't connect isbi website, I also provide isbi dataset.

wget 

• U-Net: Convolutional Networks for Biomedical Image Segmentation

• zhixuhao/unet