[Python] Improving OCR Accuracy for Digits Inside Circles from Vector PDF Images (AutoCAD...

I'm working on an OCR task where I need to detect and read numbers that are inside circles. The original data comes from .dwg (AutoCAD) files, and the PDFs I use are not scanned — they’re exported directly from AutoCAD, so they contain clean, vector-based text and shapes.

Here's my pipeline:

• 
  Convert PDF pages to images using a high resolution (up to 1200 DPI). V1.png
  
  
• 
  Detect all circles using Hough Circle Transform (they all have roughly the same radius).
  
  
• 
  Slice each region of interest (ROI) around the detected circles.
  
  
• 
  Preprocess the ROI (grayscale + thresholding).
  
  
• 
  Run Tesseract OCR on each ROI with config --psm 7 digits.
  

Despite this, I can’t seem to get the OCR accuracy above 80%. Many numbers are either misread or not detected at all, even though the input seems quite clean.

img = cv2.imread("V1.png")
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

# Detect circles
circles = cv2.HoughCircles(
gray,
cv2.HOUGH_GRADIENT,
dp=1.2,
minDist=20,
param1=50,
param2=30,
minRadius=52,
maxRadius=56
)

rod_positions = []
if circles is not None:
circles = np.round(circles[0, :]).astype("int")
for (x, y, r) in circles:
rod_positions.append((x, y, r))

rod_labels = []
for (x, y, r) in rod_positions:
patch_size_x = int(0.9 * r) - 1
patch_size_y = r // 2 - 1
roi = gray[y - patch_size_y:y + patch_size_y, x - patch_size_x:x + patch_size_x]

_, roi_thresh = cv2.threshold(roi, 150, 255, cv2.THRESH_BINARY_INV)
config = '--psm 7 digits'
number = pytesseract.image_to_string(roi_thresh, config=config)
rod_labels.append(number.strip())


What I’ve tried:

• 
  Increased PDF-to-image resolution (300 → 600 → 1200 DPI) — minor improvements.
  
  
• 
  Various Tesseract configs (--psm 6, --psm 10, etc.) — --psm 7 with digits seems best.
  
  
• 
  Different preprocessing methods (e.g., adaptive thresholding, dilation, erosion).
  

What I'm trying to achieve:

• 
  High OCR accuracy (ideally >95%) for digits inside circles.
  
  
• 
  Generate a database to train a Neural Network
  

Questions:

• 
  Given that the source is a vector PDF, should I be using a completely different approach (e.g., PDF parsing or font-based extraction)?
  
  
• 
  Are there preprocessing tricks or OCR config options that could improve digit recognition here?
  
  
• 
  Would training a custom Tesseract model help — or is that overkill for such clean, uniform data?
  

Any tips or suggestions would be hugely appreciated!

Continue reading...