图像的仿射变换原理

首先我们看看 2 维平面中姿态矩阵是如何平移、旋转和放缩物体:

from: https://www.sohu.com/a/226611009_633698

 

 

def random_rotate_xy(img, box, angle_range=(-25, 25), scale_range=(0.80, 1.0)):
    height, width = img.shape[:2]  # 获取图像的高和宽
    center = (width / 2., height / 2.)  # 取图像的中点
    angle = random.randrange(*angle_range)
    scale = random.randrange(int(scale_range[0] * 100), int(scale_range[1]) * 100) / 100.
    M = cv2.getRotationMatrix2D(center, angle, scale)  # 获得图像绕着某一点的旋转矩阵
    # cv2.warpAffine()的第二个参数是变换矩阵,第三个参数是输出图像的大小
    rotated_img = cv2.warpAffine(img, M, (width, height))
    img_h, img_w = rotated_img.shape[:2]

    # box: [x1,y1, x2,y2]
    x1, y1, x2, y2 = box
    w, h = x2 - x1, y2 - y1
    cx, cy = (x1 + x2) / 2, (y1 + y2) / 2
    cx1 = M[0, 0] * cx + M[0, 1] * cy + M[0, 2]
    cy1 = M[1, 0] * cx + M[1, 1] * cy + M[1, 2]

    x11 = max(0, int(cx1 - w * scale / 2))
    y11 = max(0, int(cy1 - h * scale / 2))
    x22 = min(img_w - 1, int(cx1 + w * scale / 2))
    y22 = min(img_h - 1, int(cy1 + h * scale / 2))

    box_rotated = [x11, y11, x22, y22]

    return rotated_img, box_rotated

 # numpy   bboxes版本

def random_rotate_with_bboxes(img, bboxes, angle_range=(-25, 25), scale_range=(0.80, 1.0)):
    """
    旋转
    :param img:
    :param bboxes: [b, 4],  format: x1y1x2y2
    :param angle: 旋转角度， >0 表示逆时针，
    :param scale:
    :return:
    """
    height, width = img.shape[:2]  # 获取图像的高和宽
    center = (width / 2, height / 2)  # 取图像的中点

    angle = random.randrange(*angle_range)
    scale = random.randrange(int(scale_range[0] * 100), int(scale_range[1]) * 100) / 100.
    M = cv2.getRotationMatrix2D(center, angle, scale)  # 获得图像绕着某一点的旋转矩阵
    # cv2.warpAffine()的第二个参数是变换矩阵,第三个参数是输出图像的大小
    rotated = cv2.warpAffine(img, M, (width, height))
    height, width = rotated.shape[:2]

    if isinstance(bboxes, np.ndarray):
        bboxes_np = bboxes
    else:
        bboxes_np = np.asarray(bboxes)

    bboxes_w, bboxes_h = bboxes_np[:, 2] - bboxes_np[:, 0], bboxes_np[:, 3] - bboxes_np[:, 1]
    bboxes_wh = np.concatenate([bboxes_w[:, None], bboxes_h[:, None]], axis=1)
    cx, cy = (bboxes_np[:, 0] + bboxes_np[:, 2]) / 2., (bboxes_np[:, 1] + bboxes_np[:, 3]) / 2.
    cxy = np.concatenate([cx[:, None], cy[:, None]], axis=1)
    M_t = np.transpose(M)
    rt_cxy = np.matmul(cxy, M_t[:2, :]) + M_t[2, :2]
    x1y1 = rt_cxy - (bboxes_wh * scale / 2)
    x2y2 = rt_cxy + (bboxes_wh * scale / 2)
    x1y1[x1y1 < 0] = 0
    x2y2[:, 0][x2y2[:, 0] > width - 1] = width
    x2y2[:, 1][x2y2[:, 1] > height - 1] = height
    bboxes_rotated = np.concatenate([x1y1, x2y2], axis=1)

    return rotated, bboxes_rotated