如何基于对象位置旋转图像？

Question

首先，对不起，这篇文章太长了

我正在做一个项目，根据叶子的图像对植物进行分类。为了减少数据的差异，我需要旋转图像，使茎在图像底部水平对齐（270度）

到目前为止我所处的位置…

到目前为止，我所做的是创建一个阈值图像，然后从中找到轮廓并围绕对象绘制一个椭圆（在许多情况下，它无法涉及整个对象，因此忽略了茎…），然后，我创建4个区域（带有椭圆的边），并尝试计算最小值区域，这是因为假设在任何一点必须找到茎，因此它将是人口较少的区域（主要是因为它将被0包围），这显然不是我想要的工作

之后，我以两种不同的方式计算旋转角度，第一种涉及atan2函数，这只需要我想从（人口最少区域的质心）移动的点，以及x=image width / 2和y = height的点。这种方法在某些情况下有效，但在大多数情况下，我没有得到所需的角度，有时需要一个负角度，它会产生一个正角度，最终茎在顶部。在其他一些情况下，它只是以一种可怕的方式失败

我的第二种方法是尝试根据3个点计算角度：图像中心、人口最少区域的重心和270º点。然后使用arccos函数，并将其结果转换为度

这两种方法对我来说都失败了

问题

• 你认为这是一个正确的方法，还是我只是把事情弄得比我应该做的更复杂
• 我如何找到叶子的茎（这不是可选的，它必须是茎）？因为我的想法不太管用
• 如何以稳健的方式确定角度？因为第二个问题中的相同原因

下面是一些示例和我得到的结果（二进制掩码）。矩形表示我正在比较的区域，穿过椭圆的红线是椭圆的长轴，粉色圆圈是最小区域内的质心，红色圆圈表示270º参考点（角度），白色圆点表示图像的中心

我当前的解决方案

    def brightness_distortion(I, mu, sigma):
        return np.sum(I*mu/sigma**2, axis=-1) / np.sum((mu/sigma)**2, axis=-1)
    
    
    def chromacity_distortion(I, mu, sigma):
        alpha = brightness_distortion(I, mu, sigma)[...,None]
        return np.sqrt(np.sum(((I - alpha * mu)/sigma)**2, axis=-1))
    
    def bwareafilt ( image ):
        image = image.astype(np.uint8)
        nb_components, output, stats, centroids = cv2.connectedComponentsWithStats(image, connectivity=4)
        sizes = stats[:, -1]
    
        max_label = 1
        max_size = sizes[1]
        for i in range(2, nb_components):
            if sizes[i] > max_size:
                max_label = i
                max_size = sizes[i]
    
        img2 = np.zeros(output.shape)
        img2[output == max_label] = 255
    
        return img2
    
    def get_thresholded_rotated(im_path):
        
        #read image
        img = cv2.imread(im_path)
        
        img = cv2.resize(img, (600, 800), interpolation = Image.BILINEAR)
        
        sat = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)[:,:,1]
        val = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)[:,:,2]
        sat = cv2.medianBlur(sat, 11)
        val = cv2.medianBlur(val, 11)
        
        #create threshold
        thresh_S = cv2.adaptiveThreshold(sat , 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY, 401, 10);
        thresh_V = cv2.adaptiveThreshold(val , 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY, 401, 10);
        
        #mean, std
        mean_S, stdev_S = cv2.meanStdDev(img, mask = 255 - thresh_S)
        mean_S = mean_S.ravel().flatten()
        stdev_S = stdev_S.ravel()
        
        #chromacity
        chrom_S = chromacity_distortion(img, mean_S, stdev_S)
        chrom255_S = cv2.normalize(chrom_S, chrom_S, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX).astype(np.uint8)[:,:,None]
        
        mean_V, stdev_V = cv2.meanStdDev(img, mask = 255 - thresh_V)
        mean_V = mean_V.ravel().flatten()
        stdev_V = stdev_V.ravel()
        chrom_V = chromacity_distortion(img, mean_V, stdev_V)
        chrom255_V = cv2.normalize(chrom_V, chrom_V, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX).astype(np.uint8)[:,:,None]
        
        #create different thresholds
        thresh2_S = cv2.adaptiveThreshold(chrom255_S , 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY, 401, 10);
        thresh2_V = cv2.adaptiveThreshold(chrom255_V , 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY, 401, 10);
            
    
        #thresholded image
        thresh = cv2.bitwise_and(thresh2_S, cv2.bitwise_not(thresh2_V))
        
        #find countours and keep max
        contours = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
        contours = contours[0] if len(contours) == 2 else contours[1]
        big_contour = max(contours, key=cv2.contourArea)
            
        # fit ellipse to leaf contours
        ellipse = cv2.fitEllipse(big_contour)
        (xc,yc), (d1,d2), angle = ellipse
        
        print('thresh shape: ', thresh.shape)
        #print(xc,yc,d1,d2,angle)
        
        rmajor = max(d1,d2)/2
        
        rminor = min(d1,d2)/2
        
        origi_angle = angle
        
        if angle > 90:
            angle = angle - 90
        else:
            angle = angle + 90
            
        #calc major axis line
        xtop = xc + math.cos(math.radians(angle))*rmajor
        ytop = yc + math.sin(math.radians(angle))*rmajor
        xbot = xc + math.cos(math.radians(angle+180))*rmajor
        ybot = yc + math.sin(math.radians(angle+180))*rmajor
        
        #calc minor axis line
        xtop_m = xc + math.cos(math.radians(origi_angle))*rminor
        ytop_m = yc + math.sin(math.radians(origi_angle))*rminor
        xbot_m = xc + math.cos(math.radians(origi_angle+180))*rminor
        ybot_m = yc + math.sin(math.radians(origi_angle+180))*rminor
        
        #determine which region is up and which is down
        if max(xtop, xbot) == xtop :
            x_tij = xtop
            y_tij = ytop
            
            x_b_tij = xbot
            y_b_tij = ybot
        else:
            x_tij = xbot
            y_tij = ybot
            
            x_b_tij = xtop
            y_b_tij = ytop
            
        
        if max(xtop_m, xbot_m) == xtop_m :
            x_tij_m = xtop_m
            y_tij_m = ytop_m
            
            x_b_tij_m = xbot_m
            y_b_tij_m = ybot_m
        else:
            x_tij_m = xbot_m
            y_tij_m = ybot_m
            
            x_b_tij_m = xtop_m
            y_b_tij_m = ytop_m
            
            
        print('-----')
        print(x_tij, y_tij)
        

        rect_size = 100
        
        """
        calculate regions of edges of major axis of ellipse
        this is done by creating a squared region of rect_size x rect_size, being the edge the center of the square
        """
        x_min_tij = int(0 if x_tij - rect_size < 0 else x_tij - rect_size)
        x_max_tij = int(thresh.shape[1]-1 if x_tij + rect_size > thresh.shape[1] else x_tij + rect_size)
        
        y_min_tij = int(0 if y_tij - rect_size < 0 else y_tij - rect_size)
        y_max_tij = int(thresh.shape[0] - 1 if y_tij + rect_size > thresh.shape[0] else y_tij + rect_size)
      
        
        x_b_min_tij = int(0 if x_b_tij - rect_size < 0 else x_b_tij - rect_size)
        x_b_max_tij = int(thresh.shape[1] - 1 if x_b_tij + rect_size > thresh.shape[1] else x_b_tij + rect_size)
        
        y_b_min_tij = int(0 if y_b_tij - rect_size < 0 else y_b_tij - rect_size)
        y_b_max_tij = int(thresh.shape[0] - 1 if y_b_tij + rect_size > thresh.shape[0] else y_b_tij + rect_size)
        
    
        sum_red_region =   np.sum(thresh[y_min_tij:y_max_tij, x_min_tij:x_max_tij])
    
        sum_yellow_region =   np.sum(thresh[y_b_min_tij:y_b_max_tij, x_b_min_tij:x_b_max_tij])
        
        
        """
        calculate regions of edges of minor axis of ellipse
        this is done by creating a squared region of rect_size x rect_size, being the edge the center of the square
        """
        x_min_tij_m = int(0 if x_tij_m - rect_size < 0 else x_tij_m - rect_size)
        x_max_tij_m = int(thresh.shape[1]-1 if x_tij_m + rect_size > thresh.shape[1] else x_tij_m + rect_size)
        
        y_min_tij_m = int(0 if y_tij_m - rect_size < 0 else y_tij_m - rect_size)
        y_max_tij_m = int(thresh.shape[0] - 1 if y_tij_m + rect_size > thresh.shape[0] else y_tij_m + rect_size)
      
        
        x_b_min_tij_m = int(0 if x_b_tij_m - rect_size < 0 else x_b_tij_m - rect_size)
        x_b_max_tij_m = int(thresh.shape[1] - 1 if x_b_tij_m + rect_size > thresh.shape[1] else x_b_tij_m + rect_size)
        
        y_b_min_tij_m = int(0 if y_b_tij_m - rect_size < 0 else y_b_tij_m - rect_size)
        y_b_max_tij_m = int(thresh.shape[0] - 1 if y_b_tij_m + rect_size > thresh.shape[0] else y_b_tij_m + rect_size)
        
        #value of the regions, the names of the variables are related to the color of the rectangles drawn at the end of the function
        sum_red_region_m =   np.sum(thresh[y_min_tij_m:y_max_tij_m, x_min_tij_m:x_max_tij_m])
    
        sum_yellow_region_m =   np.sum(thresh[y_b_min_tij_m:y_b_max_tij_m, x_b_min_tij_m:x_b_max_tij_m])
        
     
        #print(sum_red_region, sum_yellow_region, sum_red_region_m, sum_yellow_region_m)
        
        
        min_arg = np.argmin(np.array([sum_red_region, sum_yellow_region, sum_red_region_m, sum_yellow_region_m]))
        
        print('min: ', min_arg)
           
        
        if min_arg == 1: #sum_yellow_region < sum_red_region :
            
            
            left_quartile = x_b_tij < thresh.shape[0] /2 
            upper_quartile = y_b_tij < thresh.shape[1] /2
    
            center_x = x_b_min_tij + ((x_b_max_tij - x_b_min_tij) / 2)
            center_y = y_b_min_tij + (y_b_max_tij - y_b_min_tij / 2)
            
    
            center_x = x_b_min_tij + np.argmax(thresh[y_b_min_tij:y_b_max_tij, x_b_min_tij:x_b_max_tij].mean(axis=0))
            center_y = y_b_min_tij + np.argmax(thresh[y_b_min_tij:y_b_max_tij, x_b_min_tij:x_b_max_tij].mean(axis=1))
    
        elif min_arg == 0:
            
            left_quartile = x_tij < thresh.shape[0] /2 
            upper_quartile = y_tij < thresh.shape[1] /2
    
    
            center_x = x_min_tij + ((x_b_max_tij - x_b_min_tij) / 2)
            center_y = y_min_tij + ((y_b_max_tij - y_b_min_tij) / 2)
    
            
            center_x = x_min_tij + np.argmax(thresh[y_min_tij:y_max_tij, x_min_tij:x_max_tij].mean(axis=0))
            center_y = y_min_tij + np.argmax(thresh[y_min_tij:y_max_tij, x_min_tij:x_max_tij].mean(axis=1))
            
        elif min_arg == 3:
            
            
            left_quartile = x_b_tij_m < thresh.shape[0] /2 
            upper_quartile = y_b_tij_m < thresh.shape[1] /2
    
            center_x = x_b_min_tij_m + ((x_b_max_tij_m - x_b_min_tij_m) / 2)
            center_y = y_b_min_tij_m + (y_b_max_tij_m - y_b_min_tij_m / 2)
            
    
            center_x = x_b_min_tij_m + np.argmax(thresh[y_b_min_tij_m:y_b_max_tij_m, x_b_min_tij_m:x_b_max_tij_m].mean(axis=0))
            center_y = y_b_min_tij_m + np.argmax(thresh[y_b_min_tij_m:y_b_max_tij_m, x_b_min_tij_m:x_b_max_tij_m].mean(axis=1))
    
        else:
            
            left_quartile = x_tij_m < thresh.shape[0] /2 
            upper_quartile = y_tij_m < thresh.shape[1] /2
    
    
            center_x = x_min_tij_m + ((x_b_max_tij_m - x_b_min_tij_m) / 2)
            center_y = y_min_tij_m + ((y_b_max_tij_m - y_b_min_tij_m) / 2)
            
            center_x = x_min_tij_m + np.argmax(thresh[y_min_tij_m:y_max_tij_m, x_min_tij_m:x_max_tij_m].mean(axis=0))
            center_y = y_min_tij_m + np.argmax(thresh[y_min_tij_m:y_max_tij_m, x_min_tij_m:x_max_tij_m].mean(axis=1))
            
        # draw ellipse on copy of input
        result = img.copy() 
        cv2.ellipse(result, ellipse, (0,0,255), 1)

        cv2.line(result, (int(xtop),int(ytop)), (int(xbot),int(ybot)), (255, 0, 0), 1)
        cv2.circle(result, (int(xc),int(yc)), 10, (255, 255, 255), -1)
    
        cv2.circle(result, (int(center_x),int(center_y)), 10, (255, 0, 255), 5)
    
        cv2.circle(result, (int(thresh.shape[1] / 2),int(thresh.shape[0] - 1)), 10, (255, 0, 0), 5)
    
        cv2.rectangle(result,(x_min_tij,y_min_tij),(x_max_tij,y_max_tij),(255,0,0),3)
        cv2.rectangle(result,(x_b_min_tij,y_b_min_tij),(x_b_max_tij,y_b_max_tij),(255,255,0),3)
        
        cv2.rectangle(result,(x_min_tij_m,y_min_tij_m),(x_max_tij_m,y_max_tij_m),(255,0,0),3)
        cv2.rectangle(result,(x_b_min_tij_m,y_b_min_tij_m),(x_b_max_tij_m,y_b_max_tij_m),(255,255,0),3)
        
       
        plt.imshow(result)
        plt.figure()
        #rotate the image    
        rot_img = Image.fromarray(thresh)
            
        #180
        bot_point_x = int(thresh.shape[1] / 2)
        bot_point_y = int(thresh.shape[0] - 1)
        
        #poi
        poi_x = int(center_x)
        poi_y = int(center_y)
        
        #image_center
        im_center_x = int(thresh.shape[1] / 2)
        im_center_y = int(thresh.shape[0] - 1) / 2
        
        #a - adalt, b - abaix, c - dreta
        #ba = a - b
        #bc = c - a(b en realitat) 
        
        ba = np.array([im_center_x, im_center_y]) - np.array([bot_point_x, bot_point_y])
        bc = np.array([poi_x, poi_y]) - np.array([im_center_x, im_center_y])
        
        #angle 3 punts    
        cosine_angle = np.dot(ba, bc) / (np.linalg.norm(ba) * np.linalg.norm(bc))
        cos_angle = np.arccos(cosine_angle)
        
        cos_angle = np.degrees(cos_angle)
        
        print('cos angle: ', cos_angle)
        
        print('print: ', abs(poi_x- bot_point_x))
        
        m = (int(thresh.shape[1] / 2)-int(center_x) / int(thresh.shape[0] - 1)-int(center_y))
        
        ttan = math.tan(m)
        
        theta = math.atan(ttan)
            
        print('theta: ', theta) 
        
        result = Image.fromarray(result)
        
        result = result.rotate(cos_angle)
        
        plt.imshow(result)
        plt.figure()
    
        #rot_img = rot_img.rotate(origi_angle)
    
        rot_img = rot_img.rotate(cos_angle)
    
        return rot_img
    
    
    rot_img = get_thresholded_rotated(im_path)
    
    plt.imshow(rot_img)

提前谢谢 ---编辑---

我根据要求在这里留下一些原始图像。

sample

Answer 1

这就是我的意思，这需要改进。这将沿上边缘每5个像素绘制一条穿过图像中心的假想线，然后沿左边缘每5个像素绘制一条假想线，将线两侧的像素值相加，并打印最小和最大比率。元组的第四个值应该是旋转角度

from PIL import Image
import numpy as np
import math
from pprint import pprint

def rads(degs):
    return degs * math.pi / 180.0

clr = Image.open('20210210_155311.jpg').resize((640,480)).convert('L')
data = np.asarray(clr)


def ratio_lr( data, left, right ):
    x0 = left
    dx = (right-left) / data.shape[0]

    lsum = 0
    rsum = 0
    for row in range(data.shape[0]):
        lsum += data[row,:int(x0)].sum()
        rsum += data[row,int(x0):].sum()
        x0 += dx

    return lsum / rsum

def ratio_tb( data, top, bottom ):
    y0 = top
    dy = (bottom - top) / data.shape[1]

    tsum = 0
    bsum = 0
    for col in range(data.shape[1]):
        tsum += data[:int(y0),col].sum()
        bsum += data[int(y0):,col].sum()
        y0 += dy

    return tsum / bsum


midx = data.shape[1] // 2
midy = data.shape[0] // 2
track = []
for dx in range(-midx, midx, 5 ):
    if dx == 0:
        angle = 90
    else:
        angle = math.atan( midy / dx  ) * 180 / math.pi
    track.append( (ratio_lr( data, midx+dx, midx-dx ), dx, 0, angle) )

for dy in range(-midy, midy, 5 ):
    angle = math.atan( dy / midx ) * 180 / math.pi
    track.append((ratio_tb( data, midy+dy, midy-dy ), 0, dy, angle))

pprint(track)
print(min(track))
print(max(track))

Answer 2

概念

这适用于大多数叶子，只要它们有茎。因此，以下是检测旋转和旋转一个叶片图像的概念：

1. 找到叶片的近似轮廓。由于茎尖点通常属于叶片的凸面外壳（外部点），请找到轮廓的凸面外壳

2. 循环通过属于叶片凸面外壳的轮廓索引。对于每个索引，计算3个点之间的角度：索引前等高线中的点、索引处等高线中的点和索引后等高线中的点

3. 计算出的最小角度是茎尖。每次循环找到较小的角度时，将三个点存储在一个元组中，当检测到最小角度时，使用茎尖两侧的两个坐标中心和茎尖计算茎所指向的角度

4. 检测到茎的角度后，我们可以相应地旋转图像

代码

import cv2
import numpy as np

def process(img):
    img_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    img_blur = cv2.GaussianBlur(img_gray, (3, 3), 2)
    img_canny = cv2.Canny(img_blur, 127, 47)
    kernel = np.ones((5, 5))
    img_dilate = cv2.dilate(img_canny, kernel, iterations=2)
    img_erode = cv2.erode(img_dilate, kernel, iterations=1)
    return img_erode

def get_contours(img):
    contours, _ = cv2.findContours(process(img), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
    cnt = max(contours, key=cv2.contourArea)
    peri = cv2.arcLength(cnt, True)
    return cv2.approxPolyDP(cnt, 0.01 * peri, True)

def get_angle(a, b, c):
    ba, bc = a - b, c - b
    cos_angle = np.dot(ba, bc) / (np.linalg.norm(ba) * np.linalg.norm(bc))
    return np.degrees(np.arccos(cos_angle))
    
def get_rot_angle(img):
    contours = get_contours(img)
    length = len(contours)
    min_angle = 180
    for i in cv2.convexHull(contours, returnPoints=False).ravel():
        a, b, c = contours[[i - 1, i, (i + 1) % length], 0]
        angle = get_angle(a, b, c)
        if angle < min_angle:
            min_angle = angle
            pts = a, b, c
    a, b, c = pts
    return 180 - np.degrees(np.arctan2(*(np.mean((a, c), 0) - b)))

def rotate(img):
    h, w, _ = img.shape
    rot_mat = cv2.getRotationMatrix2D((w / 2, h / 2), get_rot_angle(img), 1)
    return cv2.warpAffine(img, rot_mat, (w, h), flags=cv2.INTER_LINEAR)

img = cv2.imread("leaf.jpg")
cv2.imshow("Image", rotate(img))
cv2.waitKey(0)

输出

您提供的每个示例图像的输出：

解释

分解代码：

1. 导入必要的库：

import cv2
import numpy as np

1. 定义一个函数process，将图像处理为二进制图像，使程序能够准确检测树叶的轮廓：

def process(img):
    img_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    img_blur = cv2.GaussianBlur(img_gray, (3, 3), 2)
    img_canny = cv2.Canny(img_blur, 127, 47)
    kernel = np.ones((5, 5))
    img_dilate = cv2.dilate(img_canny, kernel, iterations=2)
    img_erode = cv2.erode(img_dilate, kernel, iterations=1)
    return img_erode

1. 使用前面定义的process函数定义函数get_contours，以获得图像中最大轮廓的近似轮廓：

def get_contours(img):
    contours, _ = cv2.findContours(process(img), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
    cnt = max(contours, key=cv2.contourArea)
    peri = cv2.arcLength(cnt, True)
    return cv2.approxPolyDP(cnt, 0.01 * peri, True)

1. 定义一个函数get_angle，以获取3个点之间的角度：

def get_angle(a, b, c):
    ba, bc = a - b, c - b
    cos_angle = np.dot(ba, bc) / (np.linalg.norm(ba) * np.linalg.norm(bc))
    return np.degrees(np.arccos(cos_angle))

1. 定义函数get_rot_angle，以获取图像需要旋转的度数。它通过使用前面定义的get_angle函数，找到叶片凸包的点，其中该点与叶片轮廓中的2个周围点之间的角度最小，其中3个点之间的角度最小，从而确定该角度：

def get_rot_angle(img):
    contours = get_contours(img)
    length = len(contours)
    min_angle = 180
    for i in cv2.convexHull(contours, returnPoints=False).ravel():
        a, b, c = contours[[i - 1, i, (i + 1) % length], 0]
        angle = get_angle(a, b, c)
        if angle < min_angle:
            min_angle = angle
            pts = a, b, c
    a, b, c = pts
    return 180 - np.degrees(np.arctan2(*(np.mean((a, c), 0) - b)))

1. 使用前面定义的get_rot_angle函数，定义一个函数rotate沿图像中心旋转图像：

def rotate(img):
    h, w, _ = img.shape
    rot_mat = cv2.getRotationMatrix2D((w / 2, h / 2), get_rot_angle(img), 1)
    return cv2.warpAffine(img, rot_mat, (w, h), flags=cv2.INTER_LINEAR)

1. 最后，读入图像，应用前面定义的rotate函数并显示旋转后的图像：

img = cv2.imread("leaf.jpg")
cv2.imshow("Image", rotate(img))
cv2.waitKey(0)

Answer 3

Bilateral Symmetry

旋转图像。找到最大的轮廓。使用矩，找到该轮廓的中心。将图像分成左右两部分（注意：应用cv2.blur(img, 5,5))会产生更好的结果）：

向右翻转。覆盖左右部分：

使用cv2.absDiff（）测量左侧和（翻转）右侧之间的差异。因为叶子具有双侧对称性，当叶子的茎（或刺）垂直时，差异将最小

注：有两个最小值；一次阀杆向上，一次阀杆向下