[Python] License Plate Detection Issue - Inconsistent Results

I am working on a license plate detection system using OpenCV in Python. My script aims to detect license plates in images, apply perspective correction, and save the detected plates. However, it does not work consistently across all images. In some cases, it fails to detect plates or detects incorrect contours. Code Implementation:

Here is my current implementation:

import cv2
import os
import numpy as np
from datetime import datetime

def redresser_perspective(roi, approx):
# Function to apply perspective transformation (not included here)
pass

def detect_license_plate(image_path, output_dir='detected_plates'):
"""
Detects license plates in an image and saves them.

Enhancements:
- Perspective correction for tilted plates.
- More robust contour filtering based on aspect ratio and area.

Args:
image_path (str): Path to the input image.
output_dir (str): Directory to save detected plates.

Returns:
list: List of saved plate image paths.
"""
if not os.path.exists(output_dir):
os.makedirs(output_dir)

img = cv2.imread(image_path)
if img is None:
print(f"Error: Cannot read image {image_path}")
return []

gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

clahe = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8, 8))
enhanced = clahe.apply(gray)

blur = cv2.GaussianBlur(enhanced, (5, 5), 0)
thresh = cv2.adaptiveThreshold(blur, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY_INV, 19, 9)

rect_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
dilation = cv2.dilate(thresh, rect_kernel, iterations=1)

contours, _ = cv2.findContours(dilation, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

img_display = img.copy()
plates_saved = []
height, width, _ = img.shape

for contour in contours:
area = cv2.contourArea(contour)
peri = cv2.arcLength(contour, True)
approx = cv2.approxPolyDP(contour, 0.02 * peri, True)

if len(approx) >= 4 and len(approx) <= 8:
x, y, w, h = cv2.boundingRect(contour)

min_width = width * 0.08
min_height = height * 0.02
max_height = height * 0.2
aspect_ratio = float(w) / h
area_ratio = area / (w * h)

if (w > min_width and min_height < h < max_height and
2 < aspect_ratio < 7 and 0.3 < area_ratio < 0.95):

roi = img[y:y+h, x:x+w]

if len(approx) == 4:
roi_corrected = redresser_perspective(roi, approx)
if roi_corrected is not None:
roi = roi_corrected

roi_gray = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)
_, roi_thresh = cv2.threshold(roi_gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
white_pixels = cv2.countNonZero(roi_thresh)
total_pixels = roi_thresh.size
white_ratio = white_pixels / total_pixels

if 0.3 < white_ratio < 0.95:
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S_%f")
output_path = os.path.join(output_dir, f"plate_{timestamp}.jpg")

cv2.imwrite(output_path, roi)
plates_saved.append(output_path)

cv2.rectangle(img_display, (x, y), (x+w, y+h), (0, 255, 0), 2)

timestamp = datetime.now().strftime("%Y%m%d_%H%M%S_%f")
output_original = os.path.join(output_dir, f"original_with_plates_{timestamp}.jpg")
cv2.imwrite(output_original, img_display)

if plates_saved:
print(f"{len(plates_saved)} plate(s) detected and saved.")
else:
print("No license plate detected.")

return plates_saved


Example Images: Here are two sample images showing incorrect detections:

Original Image:



Contours Detected:



How can I improve the detection accuracy and make it more robust for different images?

Continue reading...