Image.slicer

Slice images/labels with overlapping and in different scales
Predict (with fastai) and merge the sliced images to full size

Feature

• Preprocess of high resolution imagery and label data for use in Semantic Segmentation tasks (Deep Learning)
• Increases amount of trainingdata by generating different scales and overlappings for images and labels
• Multi stage interpolation (Nearest Neighbor + Bicubic combined) for image data
• Nearest Neighbor interpolation for label data
• More than half empty slices will be ignored / It´s possilbe to slice a dismembered Mosaik!
• Add padding (to the right and bottom) to your high resolution images
• Do fastai predictions and merge the images to full size

Example

from notebook.exp.Image_Slicer import *

inp_d = Path("/Users/Abalone/Desktop/Vaihingen") 
# Directory where the Raw Images are stored
resize = 448 # Output size for the sliced Images
slice_range = [500, 900] # List of slice sizes in pixel
overlap = 7 # Ratio of overlapping (here 1:7 see image below)
dir_name = f'test4_{resize}_{overlap}_{slice_range[0]}-{slice_range[1]}' 
#dir_name = test4_448_7_500-900
#Directory name to save the outputs in 

# Dictionary to convert rgb masks to greyscale
palette = {0 : (255, 255, 255), # Impervious surfaces (white)
           1 : (0, 0, 255),     # Buildings (blue)
           2 : (0, 255, 255),   # Low vegetation (cyan)
           3 : (0, 255, 0),     # Trees (green)
           4 : (255, 255, 0),   # Cars (yellow)
           5 : (255, 0, 0),     # Clutter (red)
           6 : (0, 0, 0)}       # Undefined (black)

# Generate quantiles from slice_range (0% , 25% , 50% , 75% , 100%)
slice_l = ImageSlicer.quantile_from_slice_range(slice_range)
slice_l = [500 , 600 , 700, 800 , 900]

# Showing slices in Notebook
ImageSlicer.show_slices(slice_l, inp_d)

# Add Padding
pad = 900 # In Pixel (added to the right and bottom)
ImageSlicer.add_padding(inp_d, pad = pad)

# For Image data
ImageSlicer.slice_images(inp_d= inp_d, dir_name = dir_name, slice_l = slice_l ,
                         resize=resize, overlap = overlap)        

# For label data
inp_d_label = Path("/Users/Abalone//Test_vahingen/vaihingen_mask")
ImageSlicer.slice_masks(inp_d= inp_d2, dir_name = dir_name, slice_l = slice_l ,
                       resize= resize , overlap = overlap , palette = palette)
                       

Parameter

inp_d - Path("Path to image directory")
dir_name - Name of the directory where the images or labels are stored
resize - In pixel: What should be the dimension of the sliced Images?
slice_range - # In pixel: List of slice sizes
overlap - How much Overlap (ratio) 1:? on each side?
pad - How much Padding you want to add to your high resolution Images

optional

out_path - Path to output directory
palette - Dictionary to convert rgb masks to greyscale
slice_l - ImageSlicer.quantile_from_slice_range generates quantile from 'slice_range'
half_empty - If set to True: More than half empty slices will be ignored from the algorithm.
If set to False: No slicers will be ignored

Predict and merge

from fastai.vision import *
from PIL import Image

# Define individuell metrics from exported fastai model
def acc_isprs(input, target):
    target = target.squeeze(1)
    mask = target != void_code 
    return (input.argmax(dim=1)[mask]==target[mask]).float().mean()
metrics = acc_isprs
f_score1 = partial(fbeta , beta=1)

# Path to exported fastai model
path_model  = Path('D:/david/va_po/im_train/va_5-9_po__9-13_448')

# Type in the amount of rows and columns to rearrange the image
rows, columns = 7,7

ImageSlicer.predict_from_fastai_model(path_fastai_model= path_model,inp_d= inp_d, dir_name=dir_name, 
palette= palette , rows= rows, columns= columns )

This is a alterd version from AnmolChachra/Image_slice