我是研究生期间研究主要研究SAR影像的变化检测，这是一段简单的基于深度学习的变化检测方法，以CNN实现。
首先说下基于深度学习的变化检测任务的思路：
制作训练样本 -> 训练模型 -> 使用训练的模型遍历图片中每个像元得出结果
1 筛选训练样本
第一步，通过两幅影像生成差异图。
第二部，使用一个简单的方法对差异图进行分割，生成一幅较为粗略的二值图 。
第三部，选取变化样本和未变化样本，选择方式一般有两种：
- 1 从差异图中选择%T像素值最大和最小的像元，每个像元取其邻域分别作为变化样本和未变化样本；
- 2 从生成的粗略二值图中随机选择一部分像元，每个像元根据其类别取其邻域分别作为变化样本和未变化样本；
2 训练模型
将制作好的变化样本和未变化样本输入模型中进行训练，模型的选择通常有:
- 受限玻尔兹曼机(RBM)，使用最多
- 降噪自编码器（DAE）
- 卷积神经网络（CNN）
3 得出变化图
直接使用训练的模型遍历图片中每个像元的邻域对每个像元进行二分类即可得出检测结果。
虽然是走的监督学习的路线，但其并没有用到人工标注的gt，本质上该变化检测任务是一种非监督方法。
从深度学习的角度上讲，它属于图像分类的领域，相当于对每个像素的邻域（3×3，5×5，7×7，9×9）进行二分类。
实例代码
代码是自己写的，里面有很多地方可以完善，后续会慢慢完善，筛选样本部分耗时特别大，这里大家可自行完善一下（排序方式改一改就能快很多），或改成随机选择训练样本，注意在这个代码不更改的情况下最好只使用500×500左右的图像进行检测，太大的话会跑不动。
输入两幅影像：
import tensorflow as tf
import numpy as np
import cv2 as cv
# 读取影像
path = 'E:/data changing2/'
img1 = cv.imread(path+'1_1997.05.tif')
img2 = cv.imread(path+'1_1997.08.tif')
img1 = cv.copyMakeBorder(img1, 2, 2, 2, 2, cv.BORDER_REFLECT) #给影像填充边界
img2 = cv.copyMakeBorder(img2, 2, 2, 2, 2, cv.BORDER_REFLECT)
img_row = img1.shape[0] # 行
img_columns = img1.shape[1] # 列
img1 = np.asarray(img1, np.float32)
img2 = np.asarray(img2, np.float32)
生成差异图和差异二值图：
img_change = []
temp = np.longfloat(0)
for i in range(img_row):
for j in range(img_columns):
temp = np.square(img1[i][j][0]-img2[i][j][0]) + np.square(img1[i][j][1]-img2[i][j][1])\
+ np.square(img1[i][j][2]-img2[i][j][2])
temp = np.sqrt(temp)
img_change.append(temp)
# img_change = np.asarray(img_change, np.float32)
# max_ = img_change.max()
# min_ = img_change.min()
max_ = max(img_change)
min_ = min(img_change)
print('max = ', max_, 'min = ', min_)
for i in range(len(img_change)):
img_change[i] = (img_change[i]-min_)/(max_ - min_)*255
# 生成差异图和差异二值图
img_gray = [[0 for col in range(img_columns)] for row in range(img_row)]
img_gray = np.asarray(img_gray, np.float32)
k = 0
for i in range(img_row):
for j in range(img_columns):
img_gray[i][j] = img_change[k]
k += 1
img_gray01 = cv.imread(path + 'FCM.bmp', 0) # 使用一个比较粗的结果来筛选纯净样本
img_gray01 = cv.copyMakeBorder(img_gray01, 2, 2, 2, 2, cv.BORDER_REFLECT)
img_gray = np.asarray(img_gray, np.uint8)
img_gray01 = np.asarray(img_gray01, np.uint8)
# io.imsave(path + 'chayitu.bmp', img_gray)
# io.imsave(path + '2zhitu.bmp', img_gray01)
print("差异图，基础二值图生成完毕")
制作训练样本：
这里采用简单选择排序，邻域大小选择为7*7.
# 1. 给 img_change 排序(只排序（T%*2）以加快程序运行速度) 2. 选出(T% * 2)的样本作为训练数据集并生成相应的图像
# 1
# T 阈值
T = 0.02
T_num = int(len(img_change)*T)
print("changdu:", len(img_change))
print('T:', T_num)
# 将img_change的值赋给img_change_sort(切记不能用 = !!!! 这个会共享数据)
img_change_sort = []
for i in range(len(img_change)):
img_change_sort.append(img_change[i])
label_change = []
for i in range(len(img_change)):
label_change.append(i)
for i in range(T_num):
temp_max = img_change_sort[i]
img_label_1 = i
for j in range(i, len(img_change)):
if img_change_sort[j] > temp_max:
temp_max = img_change_sort[j]
img_label_1 = j
img_change_sort[img_label_1] = img_change_sort[i]
img_change_sort[i] = temp_max
label_change[i] = label_change[img_label_1]
label_change[img_label_1] = i
for i in range(T_num):
temp_min = img_change_sort[len(img_change_sort)-i-1]
img_label_0 = len(img_change_sort) - i - 1
for j in range(len(img_change)-i):
if img_change_sort[j] < temp_min:
temp_min = img_change_sort[j]
img_label_0 = j
img_change_sort[img_label_0] = img_change_sort[len(img_change_sort)-i-1]
img_change_sort[len(img_change_sort)-i-1] = temp_min
label_change[len(img_change_sort)-i-1] = label_change[img_label_0]
label_change[img_label_0] = len(img_change_sort) - i - 1
# 2
data = []
label = []
data1 = []
data0 = []
img_1 = [[0 for col1 in range(img_columns)] for row1 in range(img_row)]
ratio = 0.88
data_1_count = 0
data_0_count = 0
for i in range(img_row):
for j in range(img_columns):
img_1[i][j] = 125
for i in range(T_num):
region = []
row_1 = int(label_change[i]/img_columns)
column_1 = label_change[i] % img_columns
if row_1 < 2 or column_1 < 2 or row_1 >= (img_row-2) or column_1 >= (img_columns - 2):
continue
else:
count = 0
for k in range(row_1 - 2, row_1 + 3):
for l in range(column_1 - 2, column_1 + 3):
if img_gray01[k][l] == 255:
count += 1
if count / 25 >= ratio:
# if count / 49 >= 0.5:
for k in range(row_1 - 2, row_1 + 3):
for l in range(column_1 - 2, column_1 + 3):
for m in range(3):
region.append(img1[k][l][m])
for n in range(3):
region.append(img2[k][l][n])
data1.append(region)
data_1_count += 1
img_1[row_1][column_1] = 255
data1 = np.asarray(data1, np.float32)
for i in range(T_num):
row_0 = int(label_change[len(img_change)-i-1] / img_columns)
column_0 = label_change[len(img_change)-i-1] % img_columns
region = []
if row_0 < 2 or column_0 < 2 or row_0 >= (img_row-2) or column_0 >= (img_columns - 2):
continue
else:
count = 0
for k in range(row_0 - 2, row_0 + 3):
for l in range(column_0 - 2, column_0 + 3):
if img_gray01[k][l] == 0:
count += 1
if count / 25 >= ratio:
# if count / 25 >= 0.6:
for k in range(row_0 - 2, row_0 + 3):
for l in range(column_0 - 2, column_0 + 3):
for m in range(3):
region.append(img1[k][l][m])
for n in range(3):
region.append(img2[k][l][n])
data0.append(region)
# label.append(0)
data_0_count += 1
img_1[row_0][column_0] = 0
img_1 = np.asarray(img_1, np.int64)
cv.imwrite(path + 'xunlianyangben.bmp', img_1)
data0 = np.asarray(data0, np.float32)
c = min(data1.shape[0], data0.shape[0])
for i in range(c):
data.append(data1[i])
label.append(1)
data.append(data0[i])
label.append(0)
data = np.asarray(data, np.float32)
label = np.asarray(label, np.int32)
print(data_1_count, data_0_count)
print('data.shape:', np.shape(data))
# 打乱样本顺序
num_example = data.shape[0]
arr = np.arange(num_example)
np.random.shuffle(arr)
data = data[arr]
label = label[arr]
制作测试样本：
测试样本就是直接遍历整幅影像，它经过模型后输出的便是变化检测结果， 这里发现如果一个列表太大的时候会出现问题，就分成两个数组分别保存了
# 制作测试样本
data_test = []
label_test = []
for i in range(202):
for j in range(img_columns):
if i < 2 or j < 2 or i >= (img_row-2) or j >= (img_columns - 2):
continue
else:
region = []
for k in range(i - 2, i + 3):
for l in range(j - 2, j + 3):
for m in range(3):
region.append(img1[k][l][m])
for n in range(3):
region.append(img2[k][l][n])
data_test.append(region)
label_test.append(img_gray01[i][j]/255)
data_test = np.asarray(data_test, np.float32)
label_test = np.asarray(label_test, np.int32)
data_test1 = []
label_test1 = []
for i in range(202, img_row):
for j in range(img_columns):
if i < 2 or j < 2 or i >= (img_row-2) or j >= (img_columns - 2):
continue
else:
region = []
for k in range(i - 2, i + 3):
for l in range(j - 2, j + 3):
for m in range(3):
region.append(img1[k][l][m])
for n in range(3):
region.append(img2[k][l][n])
data_test1.append(region)
label_test1.append(img_gray01[i][j]/255)
data_test1 = np.asarray(data_test1, np.float32)
label_test1 = np.asarray(label_test1, np.int32)
创建并训练模型：
data = np.reshape(data, [len(data), 5, 5, 6])
data_test = np.reshape(data_test, [len(data_test), 5, 5, 6])
data_test1 = np.reshape(data_test1, [len(data_test1), 5, 5, 6])
# -----------------构建网络----------------------
# 占位符
x = tf.placeholder(tf.float32, shape=[None, 5, 5, 6], name='x')
y_ = tf.placeholder(tf.int32, shape=[None, ], name='y_')
# 第一个卷积层（5——>5)
conv1 = tf.layers.conv2d(
inputs=x,
filters=32,
kernel_size=[3, 3],
padding="same",
activation=tf.nn.relu,
kernel_initializer=tf.truncated_normal_initializer(stddev=0.01))
# 第二个卷积层(5->5)
conv2 = tf.layers.conv2d(
inputs=conv1,
filters=32,
kernel_size=[3, 3],
padding="same",
activation=tf.nn.relu,
kernel_initializer=tf.truncated_normal_initializer(stddev=0.01))
pool1 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=1)
re1 = tf.reshape(pool1, [-1, 4 * 4 * 32])
# 全连接层
logits = tf.layers.dense(inputs=re1,
units=2,
activation=tf.nn.softmax,
kernel_initializer=tf.truncated_normal_initializer(stddev=0.01),
kernel_regularizer=tf.contrib.layers.l2_regularizer(0.003))
loss = tf.losses.sparse_softmax_cross_entropy(labels=y_, logits=logits)
train_op = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
correct_prediction = tf.equal(tf.cast(tf.argmax(logits, 1), tf.int32), y_)
acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# 定义一个函数，按批次取数据
def minibatches(inputs=None, targets=None, batch_size=None, shuffle=False):
assert len(inputs) == len(targets)
if shuffle:
indices = np.arange(len(inputs))
np.random.shuffle(indices)
for start_idx in range(0, len(inputs) - batch_size + 1, batch_size):
if shuffle:
excerpt = indices[start_idx:start_idx + batch_size]
else:
excerpt = slice(start_idx, start_idx + batch_size)
yield inputs[excerpt], targets[excerpt]
# 训练CNN模型，可将n_epoch设置更大一些
n_epoch = 20
batch_size = 128
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
for epoch in range(n_epoch):
start_time = time.time()
# training
train_loss, train_acc, n_batch = 0, 0, 0
for x_train_a, y_train_a in minibatches(data, label, batch_size, shuffle=True):
_, err, ac = sess.run([train_op, loss, acc], feed_dict={x: x_train_a, y_: y_train_a})
train_loss += err
train_acc += ac
n_batch += 1
print(" train loss: %f" % (train_loss / n_batch))
# print(" train acc: %f" % (train_acc / n_batch))
将测试样本输入训练过模型得出最终结果：
# 用训练好的模型得出结果
final_img = []
final = sess.run(logits, feed_dict={x: data_test, y_: label_test})
for i in range(len(data_test)):
if final[i][0] >= final[i][1]:
finals = 0
else:
finals = 255
final_img.append(finals)
print(len(final_img))
k = 0
img_final2 = [[125 for i in range(img_columns-4)] for j in range(img_row-4)]
for i in range(200):
for j in range(img_columns-4):
img_final2[i][j] = final_img[k]
k += 1
final_img1 = []
final1 = sess.run(logits, feed_dict={x: data_test1, y_: label_test1})
for i in range(len(data_test1)):
if final1[i][0] >= final1[i][1]:
finals = 0
else:
finals = 255
final_img1.append(finals)
print(len(final_img1))
k = 0
for i in range(img_row-4-200):
for j in range(img_columns-4):
img_final2[i+200][j] = final_img1[k]
k += 1
img_final2 = np.asarray(img_final2, np.uint8)
cv.imwrite(path + 'final/2zhitu_final_not_use_CNN_5_new.bmp', img_final2)
sess.close()
完整代码: https://blog.csdn.net/weixin_39853245/article/details/97670214
