keras框架下dense_unet對稻田草地分割

如圖所示是圖像拓展後的數據集，主要分割稻田路徑。

 

老生長談，搞起來！

數據集拓展程序：

import cv2
import os
import numpy as np
from PIL import Image
# 圖像混合


# file1=r"./3dcm_png"
# file2=r"./2gu_label"
file1=r"./image"
file2=r"./mask"
files=os.listdir(file1)
files2=os.listdir(file2)


path3=r"./extendimg"
path4=r"./extendmask"
if not os.path.exists(path3):
    os.mkdir(path3)
if not os.path.exists(path4):
    os.mkdir(path4)

def contrast_img(img1, c, b):  # 亮度就是每個像素所有通道都加上b
    rows, cols, channels = img1.shape
    print(img1.shape)
    print(img1.dtype)
    # 新建全零(黑色)圖片數組:np.zeros(img1.shape, dtype=uint8)
    blank = np.zeros(img1.shape, img1.dtype)
    dst = cv2.addWeighted(img1, c, blank, 1-c, b)
    return dst
    # cv2.imshow('original_img', img)
    # cv2.imshow("contrast_img", dst)

for i in range(0,len(files)):
    path=os.path.join(file1,files[i])
    print(path)

    img1 = cv2.imread(path,cv2.IMREAD_COLOR)# (380, 308, 3) #讀取圖片1
    img1=cv2.resize(img1,(320,320),0)
    res=contrast_img(img1, 0.90, 10)
    # img1=img1.astype(np.uint8)
    # 第一幅圖的權重是 0.7 第二幅圖的權重是 0.3。大概就是第二幅圖浮現加在第一個圖上
    #g(x) = αf(x) + β，其中：α(>0)、β常稱爲增益與偏置值，分別控制圖片的對比度和亮度。
    # res = np.uint8(np.clip((0.92 * img1 + 30), 0, 255)) # 1.5是對比度   10是亮度
    # res = np.uint16(np.clip((1.15 * img1 + 80), 0, 65535)) # 1.5是對比度   10是亮度
    # res = np.uint16(np.clip((1.0 * img1 + 0.), 0, 65535)) # 1.5是對比度   10是亮度
    

    prefix=str(i+160).zfill(5)+".png"
    savepath=os.path.join(path3,prefix)
    cv2.imwrite(savepath,res) #保存
    
    # cv2.imwrite("res.jpg",res) #保存

    # cv2.imshow('dst', res)    #顯示
    # cv2.waitKey(0)
    path2m=os.path.join(file2,files2[i])
    mask=cv2.imread(path2m,0)
    print(path2m)
    mask=cv2.resize(mask,(320,320),0)
    mask=mask.astype(np.uint8)
    #mask=cv2.cvtColor(mask,cv2.COLOR_BGR2GRAY)
    mask=np.where(mask>1,255,0)
    savepath=os.path.join(path4,prefix)
    cv2.imwrite(savepath,mask) #保存

labelme轉爲實際圖片腳本：

#!/usr/bin/env python
# coding=utf-8
'''

'''
import argparse
import base64
import json
import os
import os.path as osp

import PIL.Image
import yaml

from labelme.logger import logger
from labelme import utils




NAME_LABEL_MAP = {
    '_background_': 0,
    "road": 1,
  
}



def translate(in_dir, out_dir, start_index):    
    json_files = os.listdir(in_dir)
    for jf in json_files:
        id_name = str(start_index).zfill(4)
        start_index = start_index+1
        #判斷該文件是否是JSON文件
        if jf.endswith('json'):
            json_file = os.path.join(in_dir, jf)
            with open(osp.join(out_dir, 'data.txt'), 'a') as f:                
                f.write(json_file+'\n')
            # 獲取文件名
            img_name, _ = os.path.splitext(jf)

            img_name = id_name+'_'+img_name
            
            # 獲取圖像數據
            data = json.load(open(json_file))
            imageData = data.get('imageData')
            if not imageData:
                imagePath = os.path.join(os.path.dirname(json_file), data['imagePath'])
                with open(imagePath, 'rb') as f:
                    imageData = f.read()
                    imageData = base64.b64encode(imageData).decode('utf-8')
            img = utils.img_b64_to_arr(imageData)

            # 讀取Json文件中shapes 即標註內容
            label_name_to_value = {'_background_': 0,
                                    'road': 1,
                                  
            }
            for shape in sorted(data['shapes'], key=lambda x: x['label']):
                label_name = shape['label']
                if label_name in label_name_to_value:
                    label_value = label_name_to_value[label_name]
                else:
                    label_value = len(label_name_to_value)
                    label_name_to_value[label_name] = label_value
                
                # 每一個組織存一張mask  emily
                points = shape['points']
                shape_type = shape.get('shape_type', None)
                mask_temp = utils.shape_to_mask(img.shape, points, shape_type)
                # 保存mask
                mask_temp_name=img_name+'_'+label_name+'_mask.png'
                PIL.Image.fromarray(mask_temp).save(osp.join(out_dir, mask_temp_name))



            lbl = utils.shapes_to_label(img.shape, data['shapes'], label_name_to_value)

            label_names = [None] * (max(label_name_to_value.values()) + 1)
            for name, value in label_name_to_value.items():
                label_names[value] = name
            lbl_viz = utils.draw_label(lbl, img, label_names)

            #  保存原圖
            PIL.Image.fromarray(img).save(osp.join(out_dir, img_name+'.png'))
            # 保存mask
            mask_name=img_name+'_mask.png'
            utils.lblsave(osp.join(out_dir, mask_name), lbl)
            # 保存原圖帶mask蒙版
            viz_name=img_name+'_labelviz.png'
            PIL.Image.fromarray(lbl_viz).save(osp.join(out_dir, viz_name))

            with open(osp.join(out_dir, 'label_names.txt'), 'w') as f:
                for lbl_name in label_names:
                    f.write(lbl_name + '\n')

            # logger.warning('info.yaml is being replaced by label_names.txt')
            # info = dict(label_names=label_names)
            # with open(osp.join(out_dir, 'info.yaml'), 'w') as f:
            #     yaml.safe_dump(info, f, default_flow_style=False)

            logger.info('{} Saved to: {}'.format(img_name,out_dir))




if __name__ == '__main__':
    in_dir = r'./pic'
    out_dir = r'./OUT'
    # 0 對應起始Index
    
    translate(in_dir,out_dir,0)

 下面是本章使用的denseunet分割模型

# -*- coding: utf-8 -*-
from keras.models import *
from keras.layers import Input, Concatenate, Conv2D, MaxPooling2D, UpSampling2D, Dropout, Cropping2D,BatchNormalization,Dropout,Activation,Dense
from keras.optimizers import *
from keras.callbacks import ModelCheckpoint, LearningRateScheduler
from keras import backend as keras
# from keras.regularizers import l2
from keras import regularizers
# from custom_layers import Scale
from loss import *


def DenseUnet():
    inputs = Input(shape=(320, 320, 3))

    conv1_1 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(inputs)
    BatchNorm1_1 = BatchNormalization(axis=3, gamma_regularizer=regularizers.l2(1e-4), beta_regularizer=regularizers.l2(1e-4))(conv1_1)
    ReLU1_1 = Activation('relu')(BatchNorm1_1)
    conv1_2 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(ReLU1_1)
    drop1_2 = Dropout(0)(conv1_2)
    # Merge1 = merge([conv1_1,drop1_2], mode = 'concat', concat_axis = 3)
    Merge1=Concatenate(axis=3)([conv1_1,drop1_2])
    pool1 = MaxPooling2D(pool_size=(2, 2))(Merge1)

    conv2_1 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool1)
    BatchNorm2_1 = BatchNormalization(axis=3, gamma_regularizer=regularizers.l2(1e-4), beta_regularizer=regularizers.l2(1e-4))(conv2_1)
    ReLU2_1 = Activation('relu')(BatchNorm2_1)
    conv2_2 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(ReLU2_1)
    drop2_2 = Dropout(0)(conv2_2)
    # Merge2 = merge([conv2_1,drop2_2], mode = 'concat', concat_axis = 3)
    Merge2=Concatenate(axis=3)([conv2_1,drop2_2])

    pool2 = MaxPooling2D(pool_size=(2, 2))(Merge2)

    conv3_1 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool2)
    BatchNorm3_1 = BatchNormalization(axis=3, gamma_regularizer=regularizers.l2(1e-4), beta_regularizer=regularizers.l2(1e-4))(conv3_1)
    ReLU3_1 = Activation('relu')(BatchNorm3_1)
    conv3_2 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(ReLU3_1)
    drop3_2 = Dropout(0)(conv3_2)
    # Merge3 = merge([conv3_1,drop3_2], mode = 'concat', concat_axis = 3)
    Merge3=Concatenate(axis=3)([conv3_1,drop3_2])

    pool3 = MaxPooling2D(pool_size=(2, 2))(Merge3)

    conv4_1 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool3)
    BatchNorm4_1 = BatchNormalization(axis=3, gamma_regularizer=regularizers.l2(1e-4), beta_regularizer=regularizers.l2(1e-4))(conv4_1)
    ReLU4_1 = Activation('relu')(BatchNorm4_1)
    conv4_2 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(ReLU4_1)
    drop4_2 = Dropout(0)(conv4_2)
    # Merge4 = merge([conv4_1,drop4_2], mode = 'concat', concat_axis = 3)
    Merge4=Concatenate(axis=3)([conv4_1,drop4_2])

    drop4 = Dropout(0.5)(Merge4)
    pool4 = MaxPooling2D(pool_size=(2, 2))(drop4)

    conv5_1 = Conv2D(1024, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool4)
    BatchNorm5_1 = BatchNormalization(axis=3, gamma_regularizer=regularizers.l2(1e-4), beta_regularizer=regularizers.l2(1e-4))(conv5_1)
    ReLU5_1 = Activation('relu')(BatchNorm5_1)
    conv5_2 = Conv2D(1024, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(ReLU5_1)
    drop5_2 = Dropout(0)(conv5_2)
    # Merge5 = merge([conv5_1,drop5_2], mode = 'concat', concat_axis = 3)
    Merge5=Concatenate(axis=3)([conv5_2,drop5_2])
    drop5 = Dropout(0.5)(Merge5)

    up6 = Conv2D(512, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(drop5))
    # merge6 = merge([drop4,up6], mode = 'concat', concat_axis = 3)
    merge6=Concatenate(axis=3)([drop4,up6])

    conv6_1 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge6)
    BatchNorm6_1 = BatchNormalization(axis=3, gamma_regularizer=regularizers.l2(1e-4), beta_regularizer=regularizers.l2(1e-4))(conv6_1)
    ReLU6_1 = Activation('relu')(BatchNorm6_1)
    conv6_2 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(ReLU6_1)
    drop6_2 = Dropout(0)(conv6_2)

    # Merge6 = merge([conv6_1,drop6_2], mode = 'concat', concat_axis = 3)
    Merge6=Concatenate(axis=3)([conv6_1,drop6_2])

    up7 = Conv2D(256, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(Merge6))
    # merge7 = merge([Merge3,up7], mode = 'concat', concat_axis = 3)
    merge7=Concatenate(axis=3)([Merge3,up7])

    conv7_1 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge7)
    BatchNorm7_1 = BatchNormalization(axis=3, gamma_regularizer=regularizers.l2(1e-4), beta_regularizer=regularizers.l2(1e-4))(conv7_1)
    ReLU7_1 = Activation('relu')(BatchNorm7_1)
    conv7_2 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(ReLU7_1)
    drop7_2 = Dropout(0)(conv7_2)
    # Merge7 = merge([conv7_1,drop7_2], mode = 'concat', concat_axis = 3)
    Merge7=Concatenate(axis=3)([conv7_1,drop7_2])

    up8 = Conv2D(128, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(Merge7))
    # merge8 = merge([Merge2,up8], mode = 'concat', concat_axis = 3)
    merge8=Concatenate(axis=3)([Merge2,up8])

    conv8_1 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge8)
    BatchNorm8_1 = BatchNormalization(axis=3, gamma_regularizer=regularizers.l2(1e-4), beta_regularizer=regularizers.l2(1e-4))(conv8_1)
    ReLU8_1 = Activation('relu')(BatchNorm8_1)
    conv8_2 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(ReLU8_1)
    drop8_2 = Dropout(0)(conv8_2)
    # Merge8 = merge([conv8_1,drop8_2], mode = 'concat', concat_axis = 3)
    Merge8=Concatenate(axis=3)([conv8_1,drop8_2])

    up9 = Conv2D(64, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(Merge8))
    # merge9 = merge([Merge1,up9], mode = 'concat', concat_axis = 3)
    merge9=Concatenate(axis=3)([Merge1,up9])

    conv9_1 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge9)
    BatchNorm9_1 = BatchNormalization(axis=3, gamma_regularizer=regularizers.l2(1e-4), beta_regularizer=regularizers.l2(1e-4))(conv9_1)
    ReLU9_1 = Activation('relu')(BatchNorm9_1)
    conv9_2 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(ReLU9_1)
    drop9_2 = Dropout(0)(conv9_2)
    # Merge9 = merge([conv9_1,drop9_2], mode = 'concat', concat_axis = 3)
    Merge9=Concatenate(axis=3)([conv9_1,drop9_2])

    conv9 = Conv2D(2, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(Merge9)
    conv10 = Conv2D(1, 1, activation = 'sigmoid')(conv9)#sigmoid
    #conv10 = Conv2D(1, 1, activation = 'softmax')(conv9)#sigmoid

    model = Model(input = inputs, output = conv10)
    model_dice=dice_loss(smooth=1e-5)   
    model.compile(optimizer = Adam(lr = 1e-4), loss = model_dice,metrics=['accuracy'])
    # model.compile(optimizer = Adam(lr = 1e-4), loss=[focal_loss(alpha=.25, gamma=2)],metrics=['accuracy'])
    # model.compile(optimizer = Adam(lr = 1e-4), loss = 'binary_crossentropy', metrics = ['accuracy'])

    return model

最終分割效果忘記了截圖，還是不錯的，重點是網絡模型！