M_rotate = cv2.getRotationMatrix2D((width // 2, height // 2), 45, 0.5)
print('Rotation Matrix')
print(M_rotate)
# 取得平移矩陣
M_translate = np.array([[1, 0, 100], [0, 1, -50]], dtype=np.float32)
print('Translation Matrix')
print(M_translate)
# 旋轉
img_rotate = cv2.warpAffine(img, M_rotate, (height, width))
# 平移
img_rotate_trans = cv2.warpAffine(img_rotate, M_translate, (height, width))
showImages(img=img, img_rotate=img_rotate, img_rotate_trans=img_rotate_trans)
"""
Affine Transformation - Case 2: any three point
"""
# 給定兩兩一對，共三對的點
# 這邊我們先用手動設定三對點，一般情況下會有點的資料或是透過介面手動標記三個點
height, width, _ = img.shape
points = np.array([[50, 50], [300, 100], [200, 300]], dtype=np.float32)
points_prime = np.array([[80, 80], [330, 150], [300, 300]], dtype=np.float32)
# 取得 affine 矩陣並做 affine 操作
M_affine = cv2.getAffineTransform(points, points_prime)
img_affine = cv2.warpAffine(img, M_affine, (height, width))
"""Working directory: CupoyLearning
根據以下的參考點，嘗試做透視變換
point1 = np.array([[60, 40], [420, 40], [420, 510], [60, 510]], dtype=np.float32)
point2 = np.array([[0, 80], [w, 120], [w, 430], [0, 470]], dtype=np.float32)
"""
path = "data/image/lena.png"
img = cv2.imread(path)
"""
透視轉換
"""
height, width, _ = img.shape
# 設定四對點，並取得 perspective 矩陣
w = 120
point1 = np.array([[60, 40], [420, 40], [420, 510], [60, 510]],
dtype=np.float32)
point2 = np.array([[0, 80], [w, 120], [w, 430], [0, 470]], dtype=np.float32)
# 計算 透視變換 Perspective Transformation 之矩陣
M = cv2.getPerspectiveTransform(point1, point2)
print("M\n", M)
# perspective 轉換
img_perspective = cv2.warpPerspective(img, M, (width, height))
showImages(img=img, img_perspective=img_perspective)
width = 224
height = 224
batch_size = 4
img = cv2.imread('image/Tano.JPG')
# 改變圖片尺寸
img = cv2.resize(img, (224, 224))
# cv2讀進來是BGR，轉成RGB
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img_origin = img.copy()
img = np.array(img, dtype=np.float32)
showImages(img_origin=img_origin, img=img)
# 輸入generator要是四維，(224,224,3)變成(4,224,224,3)
img_combine = np.array([img, img, img, img], dtype=np.uint8)
batch_gen = train_generator.flow(img_combine, batch_size=4)
assert next(batch_gen).shape == (batch_size, width, height, 3)
images = next(batch_gen)
images = images.astype(np.uint8)
showImages(origin=img_origin,
gen0=images[0],
gen1=images[1],
gen2=images[2],
gen3=images[3])
"""
示範如何導入 ImageDataGenerator 到 Keras 訓練中
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# region 教學
"""
模糊: 透過 Gaussian Filter 實作模糊操作
"""
img_blur = img.copy()
# 重複多次 Gaussian 模糊的操作來加深模糊的程度
img_blur1 = cv2.GaussianBlur(img_blur, (5, 5), 0)
img_blur2 = cv2.GaussianBlur(img_blur1, (5, 5), 0)
img_blur3 = cv2.GaussianBlur(img_blur2, (5, 5), 0)
showImages(img=img,
img_blur1=img_blur1,
img_blur2=img_blur2,
img_blur3=img_blur3)
"""
邊緣檢測: 透過 Sobel Filter 實作邊緣檢測
組合 x-axis, y-axis 的影像合成
"""
# 對 x 方向做 Sobel 邊緣檢測
img_sobel_x = cv2.Sobel(gray, cv2.CV_16S, dx=1, dy=0, ksize=3)
img_sobel_x = cv2.convertScaleAbs(img_sobel_x)
# 對 y 方向做 Sobel 邊緣檢測
img_sobel_y = cv2.Sobel(gray, cv2.CV_16S, dx=0, dy=1, ksize=3)
img_sobel_y = cv2.convertScaleAbs(img_sobel_y)
# x, y 方向的邊緣檢測後的圖各以一半的全重進行合成
import cv2 from utils.opencv import showImages """Working directory: CupoyLearning""" path = "data/image/lena.png" # 以彩色圖片的方式載入 img = cv2.imread(path) # 改變不同的 color space # HSL: 色相，飽和度，亮度 hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV) showImages(bgr=img, hsv=hsv)
print(model3.summary()) """ Model: "model3" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= conv2d_9 (Conv2D) (None, 576, 768, 3) 84 ================================================================= Total params: 84 Trainable params: 84 Non-trainable params: 0 _________________________________________________________________ None """ # keras 在讀取檔案實是以 batch 的方式一次讀取多張， # 但我們這裡只需要判讀一張， # 所以透過 expand_dims() 函式來多擴張一個維度 batch_img = np.expand_dims(img, axis=0) print(batch_img.shape) output1 = model1.predict(batch_img) # output1.shape = (1, 576, 768, 3) output3 = model3.predict(batch_img) # output3.shape = (1, 576, 768, 3) img1 = np.squeeze(output1, axis=0) img3 = np.squeeze(output3, axis=0) showImages(img=img, img1=img1, img3=img3)
# 建立 SIFT 物件
sift = cv2.xfeatures2d_SIFT.create()
# 偵測並計算 SIFT 特徵 (keypoints 關鍵點, descriptor 128 維敘述子)
kp_query, des_query = sift.detectAndCompute(query, None)
kp_target, des_target = sift.detectAndCompute(target, None)
"""
基於 SIFT 特徵的暴力比對
* D.Lowe ratio test
* knn 比對
"""
# 建立 Brute-Force Matching 物件
bf = cv2.BFMatcher(cv2.NORM_L2)
# 以 knn 方式暴力比對特徵
matches = bf.knnMatch(des_query, des_target, k=2)
# 透過 D.Lowe ratio test 排除不適合的配對
candidate = []
for m, n in matches:
if m.distance < 0.75 * n.distance:
candidate.append([m])
# 顯示配對結果
dst = cv2.drawMatchesKnn(query, kp_query, target, kp_target, candidate, None, flags=2)
showImages(dst=dst)
img = cv2.imread(path) # 轉為灰階圖片 gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) # 建立 SIFT 物件 """ 如果要透過 OpenCV 使用 SIFT 的話必須要額外安裝擴充的函式庫 為了避免版本問題，我們會指定安裝版本 pip install opencv-contrib-python==3.4.2.16 # 處理權限問題: ERROR: Could not install packages due to an EnvironmentError: [WinError 5] 存取被拒。 pip install --user opencv-contrib-python==3.4.2.16 """ sift = cv2.xfeatures2d.SIFT_create() # 取得 SIFT 關鍵點位置 # keypoints = sift.detect(gray, None) keypoints, features = sift.detectAndCompute(gray, None) # type(keypoints): list # #keypoints = 1098 # type(keypoints[0]): cv2.KeyPoint kp0 = keypoints[0] # 畫圖 + 顯示圖片 img_show = cv2.drawKeypoints(gray, keypoints, img.copy()) showImages(gray=gray, img_show=img_show)
""" path = "data/image/lena.png" img = cv2.imread(path) """ 上下翻轉圖片 """ # 水平翻轉 (horizontal) horizontal_flip = img[:, ::-1, :] # 垂直翻轉 (vertical) vertical_flip = img[::-1, :, :] showImages(img=img, horizontal_flip=horizontal_flip, vertical_flip=vertical_flip) """ 縮放圖片 放大 我們先透過縮小圖片去壓縮原有圖片保有的資訊，再放大比較不同方法之間的速度與圖片品質 """ # 將圖片縮小成原本的 20% small = cv2.resize(img.copy(), None, fx=0.2, fy=0.2) # 將圖片放大為"小圖片"的 8 倍大 = 原圖的 1.6 倍大 fx, fy = 8, 8 # 鄰近差值 scale + 計算花費時間
"""
# 把左上跟右下轉為矩陣型式
box = np.array((left_top, right_bottom), dtype=np.float32)
# 做矩陣乘法可以使用 `np.dot`, 為了做矩陣乘法, M_scale 需要做轉置之後才能相乘
homo_coor_result = np.dot(M.T, box)
homo_coor_result = homo_coor_result.astype('uint8')
scale_point1 = tuple(homo_coor_result[0])
scale_point2 = tuple(homo_coor_result[1])
print('origin point1={}, origin point2={}'.format(left_top, right_bottom))
print('resize point1={}, resize point2={}'.format(scale_point1, scale_point2))
# 畫圖
cv2.rectangle(small, scale_point1, scale_point2, (0, 0, 255), 3)
showImages(dst=small)
"""
Hint: 矩形
"""
img_rect = img.copy()
cv2.rectangle(
# 圖片
img_rect,
# 左上角
(60, 40),
# 右下角
(420, 510),
# 顏色
(0, 0, 255),
# 線的粗細(若為 -1 則填滿整個矩形)
lower_saturation[..., -1] = np.clip(lower_saturation[..., -1] - change_percentage, 0.0, 1.0) * 255.0
lower_saturation[..., -1] = np.int8(lower_saturation[..., -1])
# 在 HLS color space 增加飽和度
higher_saturation = hls.astype('float32').copy() / 255.0
higher_saturation[..., -1] = np.clip(higher_saturation[..., -1] + change_percentage, 0.0, 1.0) * 255.0
higher_saturation[..., -1] = np.int8(higher_saturation[..., -1])
# 轉換
lower_saturation = cv2.cvtColor(lower_saturation, cv2.COLOR_HLS2BGR)
higher_saturation = cv2.cvtColor(higher_saturation, cv2.COLOR_HLS2BGR)
# 組合圖片 + 顯示圖片
img_hls_change = np.hstack((img, lower_saturation, higher_saturation))
showImages(hls_change=img_hls_change)
"""
直方圖均衡
"""
# case 1: 把彩圖拆開對每個 channel 個別做直方圖均衡再組合起來
equal_hist = img.copy()
equal_hist[..., 0] = cv2.equalizeHist(equal_hist[..., 0])
equal_hist[..., 1] = cv2.equalizeHist(equal_hist[..., 1])
equal_hist[..., 2] = cv2.equalizeHist(equal_hist[..., 2])
showImages(img=img, equal_hist0=equal_hist)
hls_equal_hist = cv2.cvtColor(img.copy(), cv2.COLOR_BGR2HLS)
hls_equal_hist[..., 2] = cv2.equalizeHist(hls_equal_hist[..., 2])
showImages(hls=hls, hls_equal_hist=hls_equal_hist)
import cv2 from utils.opencv import showImages """Working directory: CupoyLearning""" path = "data/image/lena.png" # 以彩色圖片的方式載入 img = cv2.imread(path) # 以灰階圖片的方式載入 gray = cv2.imread(path, cv2.IMREAD_GRAYSCALE) showImages(bgr=img, gray=gray)
points = df[df.columns[:-1]].values
# 轉換為 float
points = points.astype(np.float32)
# normalize 坐標值
points = points / 96
print("圖像資料:", imgs.shape)
print("關鍵點資料:", points.shape)
# 圖像資料: (2140, 96, 96)
# 關鍵點資料: (2140, 30)
sample_img = imgs[0]
sample_points = points[0]
n_image, height, width = imgs.shape
points *= width
points = np.int(width)
n_coord = int(len(sample_points) / 2)
for i in range(n_coord):
cv2.circle(sample_img, (sample_points[i], sample_points[i + 1]),
3,
color=(0, 0, 255))
sample_img = cv2.resize(sample_img, (400, 400), interpolation=cv2.INTER_CUBIC)
showImages(sample_img=sample_img)