ARTICLE AD BOX
Problem Description
From the same input image, I compute two binary masks:
red_mask – all red-colored pixels (0 or 255)
fabric_mask – detected fabric/thread structure (0 or 255)
Both masks are correct individually and pixel-aligned.
My goal is to keep only red pixels that are structurally supported by the fabric threads, and remove red pixels that lie across gaps or outside the thread structure.
What I Expected to Happen
When combining the two masks, I expected the result to:
Preserve red pixels that lie along fabric threads
Remove red pixels that:
Fill thread gaps
Cross threads without following their geometry
Are not supported by nearby fabric pixels
In other words, the red pixels should remain only where the fabric structure exists continuously underneath them.
What Actually Happens
Using a direct logical operation:
result = cv2.bitwise_and(red_mask, fabric_mask)produces the following undesired but technically correct result:
Red pixels disappear in thin or broken fabric regions
Red pixels survive in areas where they only touch the structure at a few pixels
Red regions partially overlapping the fabric remain even if they span gaps
This happens because bitwise_and only checks per-pixel overlap, not geometric or structural consistency.
So while the operation is mathematically correct, it is semantically incorrect for this problem.
Minimal Reproducible Example
import cv2 import numpy as np img = cv2.imread("input.png") # Red mask hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV) lower_red1 = np.array([0, 70, 30]) upper_red1 = np.array([10, 255, 255]) lower_red2 = np.array([170, 70, 30]) upper_red2 = np.array([180, 255, 255]) red_mask = ( cv2.inRange(hsv, lower_red1, upper_red1) | cv2.inRange(hsv, lower_red2, upper_red2) ) # Fabric structure mask (precomputed) fabric_mask = cv2.imread("fabric_mask.png", cv2.IMREAD_GRAYSCALE) # Logical combination result = cv2.bitwise_and(red_mask, fabric_mask) cv2.imwrite("result.png", result)Why This Is Not Sufficient
Although the operation is mathematically valid, it does not encode the actual requirement:
Structural continuity
Neighborhood support
Alignment with thin, directional features
This problem is not a simple Boolean masking problem, but a structural consistency problem.
What I Am Looking For
A classical OpenCV / mathematical approach to remove red pixels that are:
Not continuously supported by the fabric structure
Only touching the structure at isolated pixels
Crossing fabric gaps or voids
Examples of acceptable approaches:
Distance-transform–based support filtering
Neighborhood-based validation
Skeleton or thickness consistency checks
Directional morphology aligned to fabric orientation
Important Note
A purely classical solution is preferred.
If this is provably not solvable using mathematical morphology alone, a lightweight model-based approach may be considered only as a last option.
Question
What is a correct classical CV approach to enforce structural support between two aligned binary masks, beyond simple pixel-wise logical operations?
