四、圖像預處理 · 數據科學和人工智能技術筆記

# 四、圖像預處理 > 作者：[Chris Albon](https://chrisalbon.com/) > > 譯者：[飛龍](https://github.com/wizardforcel) > > 協議：[CC BY-NC-SA 4.0](http://creativecommons.org/licenses/by-nc-sa/4.0/) ## 圖像二值化 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 將圖像加載為灰度 image_grey = cv2.imread('img/plane_256x256.jpg', cv2.IMREAD_GRAYSCALE) # 應用自適應閾值 max_output_value = 255 neighorhood_size = 99 subtract_from_mean = 10 image_binarized = cv2.adaptiveThreshold(image_grey, max_output_value, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, neighorhood_size, subtract_from_mean) # 展示圖像 plt.imshow(image_binarized, cmap='gray'), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/binarize_image/binarize_image_8_0.png) ## 圖像模糊 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 將圖像加載為灰度 image = cv2.imread('img/plane_256x256.jpg', cv2.IMREAD_GRAYSCALE) # 使圖像模糊 image_blurry = cv2.blur(image, (5,5)) # 展示圖像 plt.imshow(image_blurry, cmap='gray'), plt.xticks([]), plt.yticks([]) plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/blurring_img/blurring_images_8_0.png) ## 圖像剪裁 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 將圖像加載為灰度 image = cv2.imread('img/plane_256x256.jpg', cv2.IMREAD_GRAYSCALE) # 選擇所有行，和前一半的列 image_cropped = image[:,:126] # 查看圖像 plt.imshow(image_cropped, cmap='gray'), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/cropping_img/cropping_images_8_0.png) ## 邊緣檢測 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 將圖像加載為灰度 image_gray = cv2.imread('img/plane_256x256.jpg', cv2.IMREAD_GRAYSCALE) # 計算強度中值 median_intensity = np.median(image_gray) # 將閾值設為強度中值上下一個標準差 lower_threshold = int(max(0, (1.0 - 0.33) * median_intensity)) upper_threshold = int(min(255, (1.0 + 0.33) * median_intensity)) # 應用 canny 邊緣檢測 image_canny = cv2.Canny(image_gray, lower_threshold, upper_threshold) # 展示圖像 plt.imshow(image_canny, cmap='gray'), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/detect_edges/detect_edges_8_0.png) ## 增強彩色圖像的對比度 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 加載圖像 image_bgr = cv2.imread('img/plane.jpg') # 轉換為 YUV image_yuv = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2YUV) # 應用直方圖均衡 image_yuv[:, :, 0] = cv2.equalizeHist(image_yuv[:, :, 0]) # 轉換為 RGB image_rgb = cv2.cvtColor(image_yuv, cv2.COLOR_YUV2RGB) # 展示圖像 plt.imshow(image_rgb), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/enhance_contrast_of_color_image/enhance_contrast_of_color_image_12_0.png) ## 增強灰度圖像的對比度 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 將圖像加載為灰度 image = cv2.imread('img/plane_256x256.jpg', cv2.IMREAD_GRAYSCALE) # 增強圖像 image_enhanced = cv2.equalizeHist(image) # 展示圖像 plt.imshow(image_enhanced, cmap='gray'), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/enhance_contrast_of_greyscale_image/enhance_contrast_of_greyscale_image_8_0.png) # Harris 角點檢測 Harris 角點檢測器是檢測兩個邊緣角點的常用方法。它尋找窗口（也稱為鄰域或補丁），其中窗口的小移動（想象搖動窗口）使窗口內的像素內容產生大的變化。 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 將圖像加載為灰度 image_bgr = cv2.imread('img/plane_256x256.jpg') image_gray = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2GRAY) image_gray = np.float32(image_gray) # 設置角點檢測器的參數 block_size = 2 aperture = 29 free_parameter = 0.04 # 檢測角點 detector_responses = cv2.cornerHarris(image_gray, block_size, aperture, free_parameter) # 大型角點標記器 detector_responses = cv2.dilate(detector_responses, None) # 只保留大于閾值的檢測器結果，標記為白色 threshold = 0.02 image_bgr[detector_responses > threshold * detector_responses.max()] = [255,255,255] # 轉換為灰度 image_gray = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2GRAY) # 展示圖像 plt.imshow(image_gray, cmap='gray'), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/harris_corner_detector/harris_corner_detector_13_0.png) # 安裝 OpenCV 雖然有許多好的庫，OpenCV 是最受歡迎和文檔最全的圖像處理庫。使用 OpenCV 的最大障礙之一就是安裝它。但是，幸運的是，我們可以使用 Anaconda 的軟件包管理器工具 conda，在我們的終端中用一行代碼安裝 OpenCV： ``` conda install --channel https://conda.anaconda.org/menpo opencv3 ``` 之后，我們可以通過打開筆記本，導入 OpenCV 并檢查版本號（3.1.0）來檢查安裝： ```py # 加載庫 import cv2 # 查看版本號 cv2.__version__ # '3.2.0' ``` ## 顏色隔離 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 加載圖像 image_bgr = cv2.imread('img/plane_256x256.jpg') # 將 BGR 轉換為 HSV image_hsv = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2HSV) # 定義 HSV 中藍色值的范圍 lower_blue = np.array([50,100,50]) upper_blue = np.array([130,255,255]) # 創建遮罩 mask = cv2.inRange(image_hsv, lower_blue, upper_blue) # 屏蔽圖像 image_bgr_masked = cv2.bitwise_and(image_bgr, image_bgr, mask=mask) # 將 BGR 轉換為 RGB image_rgb = cv2.cvtColor(image_bgr_masked, cv2.COLOR_BGR2RGB) # 展示圖像 plt.imshow(image_rgb), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/isolate_colors/isolate_colors_14_0.png) ## 加載圖像 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 將圖像加載為灰度 image = cv2.imread('img/plane.jpg', cv2.IMREAD_GRAYSCALE) # 展示圖像 plt.imshow(image, cmap='gray'), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/load_img/load_images_4_0.png) ```py # 加載彩色圖像 image_bgr = cv2.imread('img/plane.jpg', cv2.IMREAD_COLOR) # 轉換為 RGB image_rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB) # 展示圖像 plt.imshow(image_rgb), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/load_img/load_images_6_0.png) ```py # 展示圖像數據 image ''' array([[140, 136, 146, ..., 132, 139, 134], [144, 136, 149, ..., 142, 124, 126], [152, 139, 144, ..., 121, 127, 134], ..., [156, 146, 144, ..., 157, 154, 151], [146, 150, 147, ..., 156, 158, 157], [143, 138, 147, ..., 156, 157, 157]], dtype=uint8) ''' # 展示維度 image.shape # (2270, 3600) ``` # 背景移除 [![](https://img.kancloud.cn/61/d2/61d2c3e9f7c5e7083d8e3322775c23fc_1802x1202.jpg)](https://machinelearningflashcards.com) ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 加載圖像 image_bgr = cv2.imread('img/plane_256x256.jpg') # 轉換為 RGB image_rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB) # 矩形值：起點 x，起點 y，寬度，高度 rectangle = (0, 56, 256, 150) # 創建初始遮罩 mask = np.zeros(image_rgb.shape[:2], np.uint8) # 創建用于 grabCut 的臨時數組 bgdModel = np.zeros((1, 65), np.float64) fgdModel = np.zeros((1, 65), np.float64) # 執行 grabCut cv2.grabCut(image_rgb, # 我們的圖像 mask, # 遮罩 rectangle, # 我們的矩形 bgdModel, # 用于背景的臨時數組 fgdModel, # 用于前景的臨時數組 5, # 迭代數量 cv2.GC_INIT_WITH_RECT) # 使用我們的矩形來初始化 # 創建遮罩，其中背景設置為 0，否則為 1 mask_2 = np.where((mask==2) | (mask==0), 0, 1).astype('uint8') # 使用新的遮罩移除多個圖像的背景 image_rgb_nobg = image_rgb * mask_2[:, :, np.newaxis] # 展示圖像 plt.imshow(image_rgb_nobg), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/remove_backgrounds/remove_backgrounds_13_0.png) ## 保存圖像 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 將圖像加載為灰度 image = cv2.imread('img/plane.jpg', cv2.IMREAD_GRAYSCALE) # 展示圖像 plt.imshow(image, cmap='gray'), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/save_img/save_images_4_0.png) ```py # 保存圖像 cv2.imwrite('img/plane_new.jpg', image) # True ``` ## 圖像銳化 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 將圖像加載為灰度 image = cv2.imread('img/plane_256x256.jpg', cv2.IMREAD_GRAYSCALE) # 創建核 kernel = np.array([[0, -1, 0], [-1, 5,-1], [0, -1, 0]]) # 銳化圖像 image_sharp = cv2.filter2D(image, -1, kernel) # 展示圖像 plt.imshow(image_sharp, cmap='gray'), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/sharpen_img/sharpen_images_8_0.png) ## Shi-Tomasi 角點檢測 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 加載圖像 image_bgr = cv2.imread('img/plane_256x256.jpg') image_gray = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2GRAY) # 要檢測的角點數量 corners_to_detect = 10 minimum_quality_score = 0.05 minimum_distance = 25 # 檢測角點 corners = cv2.goodFeaturesToTrack(image_gray, corners_to_detect, minimum_quality_score, minimum_distance) corners = np.float32(corners) # 在每個角點上繪制白色圓圈 for corner in corners: x, y = corner[0] cv2.circle(image_bgr, (x,y), 10, (255,255,255), -1) # 轉換為灰度 image_gray = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2GRAY) # 展示圖像 plt.imshow(image_gray, cmap='gray'), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/ski-tomasi_corner_detector/ski-tomasi_corner_detector_12_0.png) ## 使用顏色均值作為特征 ```py # 加載庫 import cv2 import numpy as np from matplotlib import pyplot as plt # 將圖像加載為 BGR image_bgr = cv2.imread('img/plane_256x256.jpg', cv2.IMREAD_COLOR) # 計算每個通道的均值 channels = cv2.mean(image_bgr) # 交換藍色和紅色值（使其變成 RGB 而不是 BGR） observation = np.array([(channels[2], channels[1], channels[0])]) # 展示通道的均值 observation # array([[ 90.53204346, 133.11735535, 169.03074646]]) # 展示圖像 plt.imshow(observation), plt.axis("off") plt.show() ``` ![png](https://chrisalbon.com/machine_learning/preprocessing_img/using_mean_color_as_a_feature/using_mean_color_as_a_feature_10_0.png)