opencv-template-matching

OpenCV Template Matching for Object Detection

Overview

OpenCV is a computer vision library that enables template matching - finding occurrences of a template image within a larger source image. This is useful for counting objects in screenshots or video frames.

Installation

pip install opencv-python numpy

Key Concepts

Template Matching

Template matching works by sliding a smaller template image across a larger image and computing a similarity score at each position. Objects are detected where the score exceeds a threshold.

Matching Methods

• cv2.TM_CCOEFF: Correlation coefficient (recommended for most cases)
• cv2.TM_CCORR: Cross correlation
• cv2.TM_SQDIFF: Sum of squared differences

Usage Examples

Convert RGB Image to Grayscale

import cv2

# Read image in color
image = cv2.imread('image.png')

# Convert to grayscale
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)

# Save grayscale image (overwrites original)
cv2.imwrite('image.png', gray)

Basic Template Matching

import cv2
import numpy as np

def count_objects(image_path, template_path, threshold=0.8):
    """Count occurrences of template in image"""
    # Read images in grayscale
    image = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
    template = cv2.imread(template_path, cv2.IMREAD_GRAYSCALE)

    # Perform template matching
    result = cv2.matchTemplate(image, template, cv2.TM_CCOEFF)

    # Find locations where correlation exceeds threshold
    locations = np.where(result >= threshold)

    # Count unique objects (accounting for nearby detections)
    count = len(locations[0])
    return count, locations

Advanced: Non-Maximum Suppression

Template matching often produces overlapping detections. Use non-maximum suppression to remove duplicates:

import cv2
import numpy as np

def count_objects_nms(image_path, template_path, threshold=0.8, min_distance=10):
    """Count objects using non-maximum suppression"""
    image = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
    template = cv2.imread(template_path, cv2.IMREAD_GRAYSCALE)

    result = cv2.matchTemplate(image, template, cv2.TM_CCOEFF)

    # Get all locations above threshold with their scores
    locations = np.where(result >= threshold)
    scores = result[locations]

    # Convert to list of (x, y, score) tuples
    detections = list(zip(locations[1], locations[0], scores))

    # Sort by score descending
    detections.sort(key=lambda x: x[2], reverse=True)

    # Apply non-maximum suppression
    kept = []
    for x, y, score in detections:
        # Check if too close to already kept detection
        too_close = False
        for kx, ky in kept:
            if abs(x - kx) < min_distance and abs(y - ky) < min_distance:
                too_close = True
                break
        if not too_close:
            kept.append((x, y))

    return len(kept), kept