Unlimblue U Net Save

u net (remote sensing image segmentation)

Project README

使用 u-net 对遥感影像进行非绿地提取

@author: sli

目录结构

加载u net

import u_net

构建网络计算图

u = u_net.u_net("cpu:0", 0, batch_size=1, output_dim=1)

添加必要的数据处理函数和可视化工具

%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np
import cv2
from libtiff import TIFF
import random
import os
from IPython.display import clear_output

def read_tif(file_path, resize=None, print_log=True):
    """
    
    参数
        file_path tif文件路径
        resize 对加载进来的图片进行resize操作,参数值为(w, h)或(w, h, c)格式的数组。该值默认为None, 表示不进行此项操作。
        print_log 是否打印图片信息 默认True
    """
    tif = TIFF.open(file_path, mode='r')
    image = tif.read_image()
    if not (resize is None):
        image = cv2.resize(image, resize)
    if print_log:
        print(image.shape)
    return image

def load_labelx(file_path, resize, max_value=1.0):
    labelx = read_tif(file_path, resize=resize)
    labelx[labelx < 128] = 0
    labelx[labelx > 128] = 255
    labelx = labelx / (255/max_value)
    return labelx

读取测试数据和真值

data = read_tif("data/test_greenland/test1.tif")
h, w, c = data.shape
label = load_labelx("data/test_greenland/testlabel1.tif", (w, h))
(3068, 3668, 3)
(3068, 3668)

显示真值图片

plt.figure(figsize=(10, 10))
plt.imshow(label)
<matplotlib.image.AxesImage at 0x7f9a14be0b38>

png

加载保存的权重,并执行分割任务

import tensorflow as tf
saver = tf.train.Saver()
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
saver.restore(sess,tf.train.latest_checkpoint('models/'))
INFO:tensorflow:Restoring parameters from models/u-net.6000-6000
p = sess.run(u.outputs, feed_dict={u.inputs: [cv2.resize(data, (572, 572))]})

分割结果保存在变量p中, 下面的函数用于给定不同阈值t对结果进行二值化处理。同时通过给定真值label,计算4个分割结果的评价指标

def tresult(label, p, t=0.35, copy=False):
    logits = p[0, :, :, 0]
    logits = cv2.resize(logits, (w, h))
    logits[logits < t] = 0.0
    logits[logits > t] = 1.0
    # TP = logits * label, TP + FP = logits, TP + FN = label, FP = logits - logits * label, TN + FP = 1 - label
    l1 = np.sum(logits * label) / np.sum(label) # Recall
    l2 = np.sum(logits * label) / np.sum(logits) # Precision
    l3 = np.sum(logits - logits * label) / np.sum(1 - label) # FPR
    l4 = l1 # TPR
    if copy:
        logits = logits.copy()
    return logits, l1, l2, l3, l4
R = []
P = []
TPR = []
FPR = []
for i in range(101):
    _, l1, l2, l3, l4 = tresult(label, p, 0.01 * i)
    R.append(l1)
    P.append(l2)
    FPR.append(l3)
    TPR.append(l4)

可视化“查全率”和“查准率”曲线

plt.plot(R, P, 'r')
plt.xlabel("Recall")
plt.ylabel("Precision")
<matplotlib.text.Text at 0x7f9a149e7198>

png

可视化“真正例率”和“假正例率”曲线

plt.plot(TPR, FPR, 'y')
plt.xlabel("TPR")
plt.ylabel("FPR")
<matplotlib.text.Text at 0x7f9a148de3c8>

png

选定一个阈值t获取一个二值化结果

logits, _, _, _, _ = tresult(label, p, 0)
print(logits.shape)
(3068, 3668)
# plt.figure(figsize=(10, 10))
plt.hist(np.ravel(logits), bins=8)
(array([ 8006314.,        0.,        0.,        0.,        0.,        0.,
               0.,  3247110.]),
 array([ 0.   ,  0.125,  0.25 ,  0.375,  0.5  ,  0.625,  0.75 ,  0.875,  1.   ]),
 <a list of 8 Patch objects>)

png

查看分类结果

plt.figure(figsize=(10, 10))
plt.imshow(logits)
<matplotlib.image.AxesImage at 0x7f99860c9ef0>

png

查看原始影像

plt.figure(figsize=(10, 10))
plt.imshow(data)
<matplotlib.image.AxesImage at 0x7f9986061e48>

png

查看分割评价图

  • 红色:分割遗漏
  • 黄色:错误分割
  • 白色:正确分割
plt.figure(figsize=(10, 10))
plt.imshow(2 * logits + label, cmap="hot")
<matplotlib.image.AxesImage at 0x7f9985148c88>

png

sess.close()
Open Source Agenda is not affiliated with "Unlimblue U Net" Project. README Source: unlimblue/u_net
Stars
37
Open Issues
0
Last Commit
6 years ago
Repository

Open Source Agenda Badge

Open Source Agenda Rating