230 - Semantic Segmentation of Landcover Dataset using U-Net

Semantic Segmentation of Landcover Dataset ​by loading images in batches from the drive​.

Code generated in the video can be downloaded from here:
https://github.com/bnsreenu/python_for_microscopists/tree/master/230_landcover_dataset_segmentation

For all code:
https://github.com/bnsreenu/python_for_microscopists

Dataset from: https://landcover.ai/
Labels:
0: Unlabeled background ​
1: Buildings​
2: Woodlands​
3: Water​

You can use any U-net but this code demonstrates the use of pre-trained encoder in the U-net - available as part of segmentation models library.

To install the segmentation models library: pip install -U segmentation-models

If you are running into generic_utils error when loading segmentation models library watch this video to fix it: https://youtu.be/syJZxDtLujs.

Prepare the data first:
1. Read large images and corresponding masks, divide them into smaller patches. And write the patches as images to the local drive.

2. Save only images and masks where masks have some decent amount of labels other than 0. Using blank images with label=0 is a waste of time and may bias the model towards unlabeled pixels.

3. Divide the sorted dataset from above into train and validation datasets.

4. You have to manually move some folders and rename them appropriately if you want to use ImageDataGenerator from keras.

After training, you can use the smooth blending process to segment large images.

Видео 230 - Semantic Segmentation of Landcover Dataset using U-Net канала DigitalSreeni