模糊检测
模糊图像特点：边缘模糊，梯度变化小。
传统方法
更多方法模糊图像检测-无参考图像的清晰度评价 - 知乎 (zhihu.com)
拉普拉斯方差
从空间域出发，分析模糊图像的梯度比较小。拉普拉斯算子测量图像的二阶导数，突出显示包含快速强度变化的图像区域。如果方差低，表示图像边缘非常少。
使用OpenCV库，技巧是设置正确阈值，阈值太低会错误将图像标记为模糊，太高，模糊图像不会被标记。
缺点：阈值需要自己设置，光照环境会对效果有很大影响。对色彩很敏感，大色块就不算模糊。无法感知运动模糊。
import cv2
image = cv2.imread(imagePath)
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
value = cv2.Laplacian(image, cv2.CV_64F).var() # 大于value表示清晰图像
快速傅里叶变化
对频域进行处理分析，模糊图像频率低；
使用matplotlib和numpy库.
缺点：对彩色图片适应良好，暗一点的图片就g。针对有清晰局部的运动模糊照片判断不行。
import numpy as np
import cv2
import imutils
def detect_blur_fft(image, size=60, thresh=10, vis=False):
# 输入： image：我们用于模糊检测的输入图像
# size：围绕图像中心点的半径大小，我们将对其进行 FFT 移位归零
# 将与幅度的平均值进行比较的值，以确定图像是否被认为是模糊的
(h, w) = image.shape
(cX, cY) = (int(w / 2.0), int(h / 2.0)) # 导出中心 (x, y) 坐标
fft = np.fft.fft2(image) #计算 FFT 以找到频率变换
fftShift = np.fft.fftshift(fft) # 然后移位零频率分量（即直流分量位于左上角）到中心
fftShift[cY - size:cY + size, cX - size:cX + size] = 0 #将 FFT 移位的中心区域清零（即，删除低频率）
fftShift = np.fft.ifftshift(fftShift) # 应用逆移使得低频再次到左上角
recon = np.fft.ifft2(fftShift) # 逆FFT
magnitude = 20 * np.log(np.abs(recon)) # 计算重建图像的幅度值，
mean = np.mean(magnitude) # 计算幅度值的平均值
return (mean, mean <= thresh) # 如果平均值小于阈值，图像将被视为“模糊”
image = cv2.imread("dataset/motion_blur2.jpg")
image = imutils.resize(image, width=500)
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
(mean, blurry) = detect_blur_fft(gray)
# print(mean, blurry)
image = np.dstack([gray] * 3)
color = (0, 0, 255) if blurry else (0, 255, 0)
text = "Blurry ({:.4f})" if blurry else "Not Blurry ({:.4f})"
text = text.format(mean)
cv2.putText(image, text, (10, 25), cv2.FONT_HERSHEY_SIMPLEX, 0.7, color, 2)
print("[INFO] {}".format(text))
# show the output image
# cv2.imshow("Output", image)
# cv2.waitKey(0)
检测视频
# import the necessary packages
from imutils.video import VideoStream
from pyimagesearch.blur_detector import detect_blur_fft
import argparse
import imutils
import time
import cv2
# construct the argument parser and parse the arguments
ap = argparse.ArgumentParser()
ap.add_argument("-t", "--thresh", type=int, default=10,
help="threshold for our blur detector to fire")
args = vars(ap.parse_args())
# initialize the video stream and allow the camera sensor to warm up
print("[INFO] starting video stream...")
vs = VideoStream(src=0).start()
time.sleep(2.0)
# loop over the frames from the video stream
while True:
# grab the frame from the threaded video stream and resize it
# to have a maximum width of 400 pixels
frame = vs.read()
frame = imutils.resize(frame, width=500)
# convert the frame to grayscale and detect blur in it
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
(mean, blurry) = detect_blur_fft(gray, size=60,
thresh=args["thresh"], vis=False)
color = (0, 0, 255) if blurry else (0, 255, 0)
text = "Blurry ({:.4f})" if blurry else "Not Blurry ({:.4f})"
text = text.format(mean)
cv2.putText(frame, text, (10, 25), cv2.FONT_HERSHEY_SIMPLEX,
0.7, color, 2)
# show the output frame
cv2.imshow("Frame", frame)
key = cv2.waitKey(1) & 0xFF
# if the `q` key was pressed, break from the loop
if key == ord("q"):
break
# do a bit of cleanup
cv2.destroyAllWindows()
vs.stop()
参考文档OpenCV Fast Fourier Transform (FFT) for blur detection in images and video streams - PyImageSearch
SVD
缺点：对色彩敏感，无法识别运动模糊。
def get_blur_degree(image_file, sv_num=10):
img = cv2.imread(image_file,cv2.IMREAD_GRAYSCALE)
u, s, v = np.linalg.svd(img)
top_sv = np.sum(s[0:sv_num])
total_sv = np.sum(s)
return top_sv/total_sv
机器学习
SVM：0-1分类；先计算数据集的算子，svm在算子之间找一个界面作为模糊和清晰界限。
缺点：动态模糊无法识别，颜色暗淡图像识别为模糊，和传统方法相比，阈值不用自己找了，但是对暗度低的图片识别效果也不行。
python代码
from sklearn.metrics import accuracy_score
from sklearn import svm
import os
import cv2
import numpy as np
import pandas as pd
from tensorflow.keras.preprocessing import image
from sklearn.metrics import accuracy_score
from sklearn.utils import shuffle
from tqdm import tqdm
# loading train data
def get_features(path):
input_size = (512, 512)
images = os.listdir(path)
features = []
for i in tqdm(images):
feature = []
# gray = cv2.imread(path+img,0)
# laplacian_var = cv2.Laplacian(gray, cv2.CV_64F).var()
img = image.load_img(path+i, target_size=input_size)
# img = image.load_img(path+i)
gray = cv2.cvtColor(np.asarray(img), cv2.COLOR_BGR2GRAY)
laplacian = cv2.Laplacian(gray, cv2.CV_64F)
feature.extend([laplacian.var(), np.amax(laplacian)])
features.append(feature)
return pd.DataFrame(features)
path_undis = '../CERTH_ImageBlurDataset/TrainingSet/Undistorted/'
path_art_blur = '../CERTH_ImageBlurDataset/TrainingSet/Artificially-Blurred/'
path_nat_blur = '../CERTH_ImageBlurDataset/TrainingSet/Naturally-Blurred/'
feature_undis = get_features(path_undis)
print('Undistorted DONE')
feature_art_blur = get_features(path_art_blur)
print('Artificially-Blurred DONE')
feature_nat_blur = get_features(path_nat_blur)
print("Naturally-Blurred DONE")
feature_art_blur.to_csv('./data/art_blur.csv', index=False)
feature_nat_blur.to_csv('./data/nat_blur.csv', index=False)
feature_undis.to_csv('./data/undis.csv', index=False)
# uncomment below code if you have pre-calculated features
# feature_art_blur = pd.read_csv('./data/art_blur.csv')
# feature_nat_blur = pd.read_csv('./data/nat_blur.csv')
# feature_undis = pd.read_csv('./data/undis.csv')
images = pd.DataFrame()
images = pd.DataFrame()
images = images.append(feature_undis)
images = images.append(feature_art_blur)
images = images.append(feature_nat_blur)
x_train = np.array(images)
y_train = np.concatenate((np.zeros((feature_undis.shape[0], )), np.ones(
(feature_art_blur.shape[0]+feature_nat_blur.shape[0], ))), axis=0)
x_train, y_train = shuffle(x_train, y_train)
# Training model
svm_model = svm.SVC(C=50, kernel='rbf')
svm_model.fit(x_train, y_train)
pred = svm_model.predict(x_train)
print('\nTraining Accuracy:', accuracy_score(y_train, pred))
CERTH数据集上87.57%准确率，我自己跑的精度87.56%。
深度学习
CNN：3层卷积。
缺点：相比支持向量机，没什么长进，可能因为层数太少了。
import numpy as np
import pandas as pd
import os
import pickle
from tensorflow.keras.preprocessing import image
import tensorflow as tf
from tensorflow.keras import utils
from sklearn.metrics import accuracy_score
input_size = (300, 300)
with open('./X_train_300.pkl', 'rb') as picklefile:
X_train = pickle.load( picklefile)
with open('./y_train_300.pkl', 'rb') as picklefile:
y_train = pickle.load( picklefile)
with open('./X_test_300.pkl', 'rb') as picklefile:
X_test = pickle.load(picklefile)
with open('./y_test_300.pkl', 'rb') as picklefile:
y_test = pickle.load(picklefile)
model = tf.keras.models.Sequential([
tf.keras.layers.Conv2D(32, (3,3), activation = 'relu', input_shape=(input_size[0], input_size[1], 3)),
tf.keras.layers.MaxPooling2D(2,2),
tf.keras.layers.Conv2D(64, (5,5), activation = 'relu'),
tf.keras.layers.MaxPooling2D(2,2),
tf.keras.layers.Conv2D(128, (5,5), activation = 'relu'),
tf.keras.layers.MaxPooling2D(2,2),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(1024, activation='relu'),
tf.keras.layers.Dropout(0.5),
tf.keras.layers.Dense(512, activation='relu'),
tf.keras.layers.Dropout(0.5),
# tf.keras.layers.Dense(256, activation='relu'),
# tf.keras.layers.Dropout(0.5),
tf.keras.layers.Dense(2, activation='softmax')
])
model.summary()
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
traindata = np.stack(X_train)
testdata = np.stack(X_test)
trainlabel = utils.to_categorical(y_train)
testlabel = utils.to_categorical(y_test)
# epochs = 10
# for i in range(epochs):
model.fit(traindata, trainlabel, batch_size=16, epochs=10, validation_data=(testdata, testlabel), verbose=1)
print("Model training complete...")
(loss, accuracy) = model.evaluate(testdata, testlabel, batch_size = 32, verbose = 1)
model.save(r'./model/model300.h5')
print("accuracy: {:.2f}%".format(accuracy * 100))
CERTH数据集上87.3%准确率，我自己跑的精度训练集73.51%。
代码链接akhileshkb/CERTH_ImageBlurDataset: model for CERTH image blur dataset challenge (github.com)
3145
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
