[Python] How to count overlapping aluminum profiles in an image with consistent cross-section?...

I have an image showing several aluminum profile cross-sections, all with a consistent shape. However, due to overlapping, tilting, and plastic wrapping, it’s difficult to count them using traditional method



import cv2
import numpy as np
import matplotlib.pyplot as plt

# Load image
image = cv2.imread("aluminum_profiles.png")
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)

# Edge detection
edges = cv2.Canny(gray, 50, 150)

# Morphological operations to clean up
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))
cleaned = cv2.morphologyEx(edges, cv2.MORPH_CLOSE, kernel)

# Find contours
contours, _ = cv2.findContours(cleaned, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

# Filter and count large contours
count = 0
for cnt in contours:
area = cv2.contourArea(cnt)
if area > 1000: # Empirical threshold
count += 1
cv2.drawContours(image, [cnt], -1, (0, 255, 0), 2)

print("Detected profiles:", count)

# Show result
plt.imshow(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))
plt.axis('off')
plt.show()

1. 
   I tried OpenCV Canny edge detection and findContours(). However, overlapping shapes and packaging noise make it hard to isolate individual profiles.
   
   
2. 
   Using one aluminum profile as a template in cv2.matchTemplate() struggles when the shapes are rotated or partially occluded, and results in many false positives.
   
   
3. 
   With SIFT/ORB, matching keypoints were scattered, making it difficult to group them into distinct instances.
   

All the above methods either overcount or fail to resolve individual profiles under occlusion and tilt.

I want a robust method to count the aluminum profiles even under:
• Partial occlusion.
• Small rotations.
• Noise (e.g., plastic wrapping).

What would be a better approach for this type of repeated but imperfectly aligned object counting?

Should I try LoFTR / SuperGlue for feature matching?

Would training a lightweight instance segmentation model like YOLOv8 / Mask R-CNN be more effective?

Or is there a traditional CV trick I’m missing?

Appreciate any suggestions, examples, or implementation tips!

Continue reading...