使用 OpenCV 计算昆虫中的重叠对象问题

Question

我正在数一群蟋蟀（昆虫）的数量。我将使用 openCV 库的图像处理方法。当农民出售他们的蟋蟀时，这将提供更准确的练习。照片是用智能手机拍摄的。不幸的是，结果并不如预期。由于大多数蟋蟀相互重叠，我的代码无法将它们分开，导致计数不正确。

我应该用什么方法来解决这个问题？我的代码有问题吗？

蟋蟀图片

这是我的代码。

import cv2
import numpy as np

img = cv2.imread("c1.jpg",1)

roi=img[0:1500,0:1100]  
gray = cv2.cvtColor(roi,cv2.COLOR_BGR2GRAY)
gray_blur=cv2.GaussianBlur(gray,(15,15),0)
thresh = cv2.adaptiveThreshold(gray_blur,255,cv2.ADAPTIVE_THRESH_GAUSSIAN_C,cv2.THRESH_BINARY_INV,11,1)
kernel=np.ones((1,1),np.uint8)
closing=cv2.morphologyEx(thresh,cv2.MORPH_CLOSE,kernel,iterations=10)

result_img=closing.copy()
contours,hierachy=cv2.findContours(result_img,cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_SIMPLE)
counter=0
for cnt in contours:
    area = cv2.contourArea(cnt)
    if area <  150  :
    #if area<  300 :
        continue
    counter+=1
    ellipse = cv2.fitEllipse(cnt)
    cv2.ellipse(roi,ellipse,(0,255,0),1)
        
  
cv2.putText(roi,"Crickets="+str(counter),(100,70),cv2.FONT_HERSHEY_SIMPLEX,1,(255,0,0),1,cv2.LINE_AA)
cv2.imshow('ImageOfCrickets',roi)
#cv2.imshow('ImageOfGray',gray)
#cv2.imshow('ImageOfGray_blur',gray_blur)
#cv2.imshow('ImageOfThreshold',thresh)
#cv2.imshow('ImageOfMorphology',closing)

print('Crickets = '+ str(counter))     

cv2.waitKey(0)
cv2.destroyAllWindows()

现在，我正在使用闭合形态学和等高线层次结构进行椭圆形方法。

Answer 1

这里有一个选项。使用自适应阈值，执行腐蚀/扩张和高斯模糊，然后轮廓，然后按大小和纵横比过滤它们，最后找到每个过滤轮廓的质心。

import cv2

# Load the image
img = cv2.imread('insects.jpg')

# Convert to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
gray = cv2.GaussianBlur(gray,(3,3),2)
# Threshold the grayscale image
thresh = cv2.adaptiveThreshold(gray,300,cv2.ADAPTIVE_THRESH_GAUSSIAN_C,cv2.THRESH_BINARY_INV,85,-21)

# # Perform morphological operations
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (9,9))
thresh = cv2.erode(thresh, kernel, iterations=2)
thresh = cv2.dilate(thresh, kernel, iterations=1)
thresh = cv2.GaussianBlur(thresh, (3,3), 1)

# Find contours
contours, hierarchy = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

print(contours)

# Filter contours by area and aspect ratio
min_area = 500  # minimum area of contour
max_area = 10000  # maximum area of contour
min_aspect_ratio = 0.3  # minimum aspect ratio of contour
max_aspect_ratio = 3  # maximum aspect ratio of contour

filtered_contours = []
for contour in contours:
    area = cv2.contourArea(contour)
    x, y, w, h = cv2.boundingRect(contour)
    aspect_ratio = float(w) / h if h != 0 else 0
    if area >= min_area and area <= max_area and aspect_ratio >= min_aspect_ratio and aspect_ratio <= max_aspect_ratio:
        filtered_contours.append(contour)
        print(contour)

# Compute centers of mass and draw circles for filtered contours
for contour in filtered_contours:
    # Compute moments of the contour
    M = cv2.moments(contour)
    if M['m00'] != 0:
        # Compute center of mass
        cx = int(M['m10'] / M['m00'])
        cy = int(M['m01'] / M['m00'])
        # Draw circle at center of mass
        cv2.circle(img, (cx, cy), 5, (0, 255, 0), -1)


# Show the original image with filtered contours
cv2.drawContours(img, filtered_contours, -1, (0, 0, 255), 2)
cv2.imwrite('Image_contours.jpg', img)
cv2.imshow('Gray', gray)
cv2.imwrite('image_thresholded_preprocessed.jpg', thresh)
cv2.waitKey(0)
cv2.destroyAllWindows()

另一种选择，类似于这里所做的事情：

• 使用 OpenCV 和 Python 检测接触/重叠的圆圈/椭圆。这是在那次的启发。

图像是经过阈值处理的（不是昆虫是白色的），以及带有中心（红点）的轮廓。在现实中不是很好，但我能想到的最好的。似乎有很多工作要做，但解决误报（在空白处）的方法可能是通过第二次传递去除质心不暗的轮廓。

Answer 2

这是一个答案，使用新发布的 Meta 的

Segment Anything

库来完成繁重的工作，而 OpenCV 仅用于操作。我刚看了一个关于 Segment Anything 的 youtube 视频，想用这个问题来测试一下，这个问题困扰了我一段时间——我喜欢这个例子，因为对我来说这并不容易。

roboflow 的这篇博文非常有用：Roboflow：如何使用分段任何模型 (SAM)。

github页面中的信息也很有帮助：GitHub Facebook Research: Automatic mask generation (facebook research's Segment Anything)

这是代码，最后计算出 135 只昆虫（一些重复计算）。为了提取适当的蒙版，蒙版按区域（相对于图像区域）进行过滤。可能有更好的过滤方法，例如，您还可以按纵横比过滤 - 这可能会丢弃腹部并允许增加区域范围。

这是代码。很短，但有点慢。也许使用灰度可以加快速度。

# Import Libraries
import cv2, torch
import supervision as sv
from segment_anything import SamAutomaticMaskGenerator, sam_model_registry, SamAutomaticMaskGenerator

# Note that Segment anything requires that you download the checkpoint.  In my case I downloaded it to the same path as the image

# Replace the following with the actual paths and model type
model_type = "default" # default is the same as vit_h
checkpoint_path = "sam_vit_h_4b8939.pth" # you have to download this from the github 
image_path = "insects.jpg"
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
min_area = 0.1/100  # 0.1% of image area (insects should be larger than 5% of the image area)
max_area = 0.6/100 # 0.6% of image area (insects should be smaller than 10% of the image area)


# Print Information for this run
print('Device in use =', DEVICE)

# Load the image
image = cv2.imread(image_path)
image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

# Create the SAM model and mask generator
sam = sam_model_registry[model_type](checkpoint=checkpoint_path)
mask_generator = SamAutomaticMaskGenerator(sam)

# Generate masks for the entire image
masks = mask_generator.generate(image_rgb)

# Calculate the threshold for mask area (10% of the image size)
area_hi_thres = max_area * image.shape[0] * image.shape[1]
area_low_thres = min_area * image.shape[0] * image.shape[1]

# Filter the masks based on their area
filtered_masks = [mask for mask in masks if area_low_thres < mask['area'] < area_hi_thres]

# Use supervision package to anotate masks
mask_annotator = sv.MaskAnnotator()
detections = sv.Detections.from_sam(filtered_masks)
annotated_image = mask_annotator.annotate(image, detections)

# Write Images to File
cv2.imwrite('Image RGB.jpg', image_rgb) # it will likely work also in gray scale and a bit faster
cv2.imwrite('annoated_image.jpg', annotated_image)

# Print the count
print('Number of Objects', len(filtered_masks))