[Python] Robustly contour freeform paper in OpenCV under unreliable & variable lighting?

I'm building a vision system in OpenCV that must robustly detect a sheet of paper under varying lighting and real-world conditions like shadows, curves, wrikles...

The goal is to create an algorithm that:

• Detects the full sheet of paper even if curved
• Rejects spike offshoots (e.g., false edges due to shadows or texture - basically where the contour gets all crazy).
• Bridges gaps where the contour offshoots

Basically, to get the contour to close the loop around the paper rather than sprawling out all over the wall.

My contour code


def Contour():
img = cv2.imread("imgage.jpg")
preprocessed = Preprocess(img)

# Find contours
contours = cv2.findContours(preprocessed, cv2.RETR_EXTERNAL, cv2.CHAIN_APPR$
contours = contours[0] if len(contours) == 2 else contours[1]

# Take the largest contour only
if contours:
largest = max(contours, key=cv2.contourArea)
cv2.drawContours(img, [largest], -1, (0, 255, 0), 2)
else:
print("No contours found.")

# Show result
cv2.imshow("Result", img)
cv2.waitKey(0)
cv2.destroyAllWindows()



I also tried polygon approximation, but that didn't really solve anything (I tried a lot of values for epsilon):

epsilon = 0.001 * cv2.arcLength(largest, True)
poly = cv2.approxPolyDP(largest, epsilon, True)




I could post a ton more pictures but the idea is the same - an algorithm that corrects this is needed not just improvements to the preprocessing code (as shown in bonus) because it gets worse and changes nature with different lighting setups (ideally, if a human can contour it, so should it).

BONUS

BONUS 1: Sometimes the top of the paper is occluded by binder clips. The contour goes around them currently. How can I "connect the dots" across these missing sections and replace with one straight line in a plausible way?


BONUS 2: Is there a better way to preprocess that can help minimize the risk of offshoots from shadows or noise? I still need the algorithm above for robustness, but improved preprocessing could help reinforce it.
​

Please find my preprocessing code below:

def Preprocess(img):
# Grayscale image
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

# Local contrast enhancement
clahe = cv2.createCLAHE(clipLimit=2, tileGridSize=(48, 48))
clahe_img = clahe.apply(gray)

# Blur to reduce noise
blurred = cv2.GaussianBlur(clahe_img, (7, 7), 0)

# Adaptive threshold to isolate white regions
thresh = cv2.adaptiveThreshold(blurred, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY, 51, 2)
return thresh


Note: We cannot change our hardware/setup and are not looking for any recommendations on that front

Continue reading...