1. Dataset download

image image
mask image
label image

Download link
Extraction code: 2022
The sample image and label are stored in the dataset folder, and you can view their channels and other information to facilitate processing your own dataset.

2. Dataset preprocessing

Refer to another article: Image preprocessing

Dataset image
1. is 24 bit depth
2. 3-channel RGB image
Dataset label
1. is 8 bit depth
2. The value of the label is 0, 1, 2, ... (n-1 if there are n types of values)

If the value of your data set label is 255, it needs to be converted to 0, 1, 2, etc. ( depending on your own situation )
For example:

Background [0,0,0]--------------0

Person [192,128,128]--------------1

Bike [0,128,0]----------------------2

Car [128,128,128]----------------- 3

Drone [128,0,0]--------------------4

Boat [0,0,128]--------------------- 5

Animal [192,0,128]---------------- 6

Obstacle [192,0,0]------------------7

Construction [192,128,0]-----------8

Vegetation [0,64,0]-----------------9

Road [128,128,0]-------------------10

Sky [0,128,128]---------------------11

If the video memory is not enough, you can crop the image to a size of 256, the corresponding code is as follows:
the cropped image is saved in image and label

import os
import numpy as np
import cv2

images_path = './JPEGImages/'
labels_path = './SegmentationClass/'

image_files = os.listdir(images_path)

for s in image_files:
    image_path = images_path + s
    label_path = labels_path + s[:-4]+'.png'
    image = cv2.imread(image_path)
    label = cv2.imread(label_path)
    #print(image.shape)
    index = 0
    for i in range(4):
        for j in range(4):
            #print(i*256, ((i+1)*256-1), j*256, ((j+1)*256-1))
            new_image = image[i*256 : ((i+1)*256), j*256 : ((j+1)*256), :]
            new_label = label[i*256 : ((i+1)*256), j*256 : ((j+1)*256), :]
            cv2.imwrite('./image/'+ 'b_' + s[:-4] +'_'+ str(index) + '.png', new_image)
            new_label = cv2.cvtColor(new_label, cv2.COLOR_BGR2GRAY)
            cv2.imwrite('./label/'+ 'b_' + s[:-4] +'_'+ str(index) + '.png', new_label)
            index+=1
    print(s)

3. Model download

download link

4. Check the data

After the data preparation is complete, first run voc_annotation.pythe file, generate a data list, and run the command: python voc_annotation.py, generally no error will be reported here, if an error is reported, you can check it in the following way:
1. Check if the label is 8位深度, if not, run the following code:
```
import os

import cv2

file_names = os.listdir('./SegmentationClass/')

for s in file_names:
    image_path = os.path.join('./SegmentationClass/', s)
    image = cv2.imread(image_path)
    image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    cv2.imwrite('./SegmentationClass/' + s , image)
```
2. Check if there are other values in the label. For example 类别为3, the value in the label can only be 0，1，2. If there are other values, an error may be reported; even if there is no error reported here, there may be problems in training, so if there is this problem , must remember to modify. If there are not too many such values, you can directly set them to 0; if they are too many, you need to check which step caused the problem. If you can’t find it, it is recommended to change the data set.

5. Train the model

Modify the train.py code
image size (important)

image category (important)

batch_size and other parameters as the case may be
run python train.pyto train

6. Model Evaluation

Modify the content of deeplab.py, which are the trained models. Number of categories (including background). image size
Modify get_miou.py content
run python get_miou.pycommand

7. Model Prediction

Modify the content of deeplab.py, which are the trained models. Number of categories (including background). image size
Modify predict.py content. name_classes indicates the name corresponding to each category
run python predict.pycommand

deeplab implements its own remote sensing geological segmentation data set