实验——DISK

DISK提取网络的鲁棒性分析

原始图像提取特征

均值滤波

1.对少量图片作均值滤波处理

import cv2
import os

data_path = '/home/julia/test_images'
save_path = '/home/julia/images_gauss'

if not os.path.exists(save_path):
    os.mkdir(save_path)
img_list = os.listdir(data_path)
for i in range(len(img_list)):
    print(i)
    img_path = os.path.join(data_path, img_list[i])
    image = cv2.imread(img_path, cv2.IMREAD_COLOR)
    # gaussian
    image = cv2.blur(image, (5, 5))
    cv2.imwrite(save_path+'/'+str(i+1)+'.jpg', image)

2.运行detect.py
python detect.py h5_artifacts_destination /home/julia/images_gauss --height=1680 --width=1280

3.运行view.py
python view_h5.py h5_artifacts_destination /home/julia/images_gauss keypoints

4.查看提取结果


5.结论
点的提取没有原图密集，但鲁棒性增强

光照

1.对少量图片作中心光照处理

import cv2
import os
import math
import numpy as np
data_path = '/home/julia/test_images'
save_path = '/home/julia/images_light'

if not os.path.exists(save_path):
    os.mkdir(save_path)
img_list = os.listdir(data_path)
for i in range(len(img_list)):
    print(i)
    img_path = os.path.join(data_path, img_list[i])
    image = cv2.imread(img_path, cv2.IMREAD_COLOR)
    # light
    # 获取图像行和列
    rows, cols = image.shape[:2]
    # 设置中心点
    centerX = rows / 2
    centerY = cols / 2
    radius = min(centerX, centerY)
    # 设置光照强度
    strength = 200
    # 图像光照特效
    for i in range(rows):
        for j in range(cols):
            # 计算当前点到光照中心距离(平面坐标系中两点之间的距离)
            distance = math.pow((centerY - j), 2) + math.pow((centerX - i), 2)
            # 获取原始图像
            B = image[i, j][0]
            G = image[i, j][1]
            R = image[i, j][2]
            if (distance < radius * radius):
                # 按照距离大小计算增强的光照值
                result = (int)(strength * (1.0 - math.sqrt(distance) / radius))
                B = image[i, j][0] + result
                G = image[i, j][1] + result
                R = image[i, j][2] + result
                # 判断边界 防止越界
                B = min(255, max(0, B))
                G = min(255, max(0, G))
                R = min(255, max(0, R))
                image[i, j] = np.uint8((B, G, R))
            else:
                image[i, j] = np.uint8((B, G, R))
    cv2.imwrite(save_path+'/'+str(i+1)+'.jpg', image)

2.运行detect.py
python detect.py h5_artifacts_destination /home/julia/images_light --height=1680 --width=1280

3.运行view.py
python view_h5.py h5_artifacts_destination /home/julia/images_light keypoints

4.查看提取结果


5.结论
过强的光照导致图片失真，也会导致提取网络性能降低

高斯噪声
1.对少量图片作高斯噪声处理

import numpy as np
import cv2
import os

# add gauss_noise to image
def gasuss_noise(image, mean=0, var=0.005):
    # change image to array
    image = np.array(image/255, dtype=float)
    # creat noise
    noise = np.random.normal(mean, pow(var, 0.5), image.shape)
    # use noise
    out = image + noise
    if out.min() < 0:
        low_clip = -1
    else:
        low_clip = 0
    out = np.clip(out, low_clip, 1.0)
    out = np.uint8(out*255)
    return out

old_path = '/home/julia/images_test'
new_path = '/home/julia/images_gaussloss'

# if not exist make dir
if not os.path.exists(new_path):
    os.mkdir(new_path)
# get the image list
img_list = os.listdir(old_path)
for i in range(len(img_list)):
    print(i)
    # get the path of every image
    old_image_path = os.path.join(old_path, img_list[i])
    img = cv2.imread(old_image_path, cv2.IMREAD_UNCHANGED)
    rows, cols, dims = img.shape
    img = gasuss_noise(img)
    # save the new image
    new_image_path = os.path.join(new_path, img_list[i])
    cv2.imwrite(new_image_path, img, [int(cv2.IMWRITE_JPEG_QUALITY), 100])

2.运行detect.py
python detect.py h5_artifacts_destination /home/julia/images_gaussloss --height=1680 --width=1280

3.运行view.py
python view_h5.py h5_artifacts_destination /home/julia/images_gaussloss keypoints

4.查看提取结果


5.结论
提取网络不能很好地处理带高斯噪声的图片

椒盐噪声
1.对少量图片作椒盐噪声处理

import numpy as np
import cv2
import os


# add sp_noise to image
def sp_noise(img, rows, cols):
    # creat white point
    for j in range(10000):
        x = np.random.randint(0, rows)
        y = np.random.randint(0, cols)
        img[x, y, :] = 255
    # creat black point
    for k in range(10000):
        x = np.random.randint(0, rows)
        y = np.random.randint(0, cols)
        img[x, y, :] = 0
    return img
    
old_path = '/home/julia/images_test'
new_path = '/home/julia/images_gaussloss'

# if not exist make dir
if not os.path.exists(new_path):
    os.mkdir(new_path)
# get the image list
img_list = os.listdir(old_path)
for i in range(len(img_list)):
    print(i)
    # get the path of every image
    old_image_path = os.path.join(old_path, img_list[i])
    img = cv2.imread(old_image_path, cv2.IMREAD_UNCHANGED)
    rows, cols, dims = img.shape
    img = sp_noise(img, rows, cols)
    # save the new image
    new_image_path = os.path.join(new_path, img_list[i])
    cv2.imwrite(new_image_path, img, [int(cv2.IMWRITE_JPEG_QUALITY), 100])

2.运行detect.py
python detect.py h5_artifacts_destination /home/julia/images_sploss --height=1680 --width=1280

3.运行view.py
python view_h5.py h5_artifacts_destination /home/julia/images_sploss keypoints

4.查看提取结果

5.结论
提取网络不能很好地处理带椒盐噪声的图片