Build your own semantic segmentation platform deeplabv3+

---

Preface

In the last article, I mainly learned about semantic segmentation , Instance segmentation , Difference of panoramic segmentation , as well as labelme Convert the marked data , This article is mainly through deeplabV3+ Build your own semantic segmentation platform

One 、deeplabV3+

The picture above , yes deeplabV3+ The main frame of , Simply put, it's coding , Decoding process . Pass the input picture through DCNN Deep convolution neural network , Get two effective feature layers （ Shallow ）（ Deep level ） The deep characteristic layer ASPP（ The expansion convolution with different expansion rates is used for feature extraction , Then stack the features , adopt 1X1 Convolution adjusts the number of channels , Get the final feature ） High semantic feature information is combined with shallow features through up sampling , It's going on 3X3 Convolution of , And then through 1*1 Convolution adjusts the number of channels , Adjust to num_class（ Number of categories ） Upsampling is performed so that the final output layer , The width and height are the same as the input image , Get every kind of every pixel .

Two 、 Data preparation

1. First of all, we need to deal with the data

JPEGImages It stores pictures

SegmentationClass Deposit is mask Mask image

ImageSets Storage is some txt file

3、 ... and 、 Modify the code

1. In the root directory mypath.py file

2.dataloaders\datasets Create your own dataset file hat.py
Copy the pascal.py file

3.dataloaders/utils.py


4.dataloaders/__init__.py


5.train.py

Four 、 Start training

Some main parameters



Then you can train directly ：

You can also carry a server for training , You can read my previous article .

5、 ... and 、 test

After training , So we can test it , Let's take a look at the code .

import torch
from modeling.deeplab_v3_50_modify_256 import deeplab_v3_50
import glob
import cv2
import os
from modeling.deeplab import *

from PIL import Image
from torchvision import transforms
from dataloaders.utils import  *
from utils.saver import save_colored_mask
num_class=3

path = 'test_image'
out_path='out_image'

test_images = glob.glob(os.path.join(path,"*.jpg"))

composed_transforms = transforms.Compose([transforms.ToTensor()])
totensor = transforms.ToTensor()

model=DeepLab(num_classes=num_class,backbone='drn')
model.load_state_dict(torch.load(r'D:\xiangmu\deeplaV3_run\run_hat\hat\deeplab-drn\model_best.pth.tar')['state_dict'])
model.eval()

def Normalize(img,mean,std):
    img = np.array(img).astype(np.float32)
    img /= 255.0
    img -= mean
    img /=std
    return img


for test_image in test_images:
    name=os.path.basename(test_image)
    name=name.replace('jpg','png')


    img = Image.open(test_image)
    img_norm = Normalize(img,mean=(0.485, 0.456, 0.406),std=(0.229, 0.224, 0.225))
    img_resize = cv2.resize(img_norm,(513,513))
    compose = composed_transforms(img_resize).unsqueeze(0)
    out = model(compose)
    pred=torch.argmax(out,1)[0].numpy()
    H,W=img_norm.shape[0],img_norm.shape[1]
    pred_orgin=cv2.resize(pred,(W,H),interpolation=cv2.INTER_NEAREST)
    out_file=os.path.join(out_path,name)
    save_colored_mask(pred_orgin,out_file)
    print('save {}  Testing is completed '.format(out_file))

Test results ：