Edges and contours

The idea#

Where an image jumps from light to dark there is an edge. A computer finds those places by subtracting neighbouring pixels: a small difference means a surface, a large one means an edge.

What it is good for#

• Counting parts on a conveyor.
• Checking whether a component has the right outline.
• Finding a document in a photo and rectifying it.
• Measuring lines, circles and rectangles.

Canny with automatic thresholds#

import cv2, numpy as np

def auto_canny(grey, sigma=0.33):
    v = float(np.median(grey))
    lower = int(max(0, (1.0 - sigma) * v))
    upper = int(min(255, (1.0 + sigma) * v))
    return cv2.Canny(grey, lower, upper)

img  = cv2.imread("parts.jpg")
grey = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
grey = cv2.GaussianBlur(grey, (5, 5), 0)     # without this: noise edges
edges = auto_canny(grey)

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

for c in contours:
    area = cv2.contourArea(c)
    if area < 800:
        continue
    perim = cv2.arcLength(c, True)
    corners = cv2.approxPolyDP(c, 0.02 * perim, True)
    circ = 4 * np.pi * area / (perim ** 2)         # 1.0 = perfect circle
    shape = "circle" if circ > 0.85 else f"{len(corners)}-gon"
    print(f"area {area:8.0f}  circularity {circ:.2f}  -> {shape}")

The contour features you need#

The gradient#

Canny's four steps#

1. 01

   Smooth

   Gaussian filter, so that noise does not create gradients.

2. 02

   Compute the gradient

   Sobel in both directions, then magnitude and direction.

3. 03

   Non-maximum suppression

   Only the strongest point across the edge direction survives, which makes the lines one pixel thin.

4. 04

   Hysteresis

   Points above the upper threshold are certainly edges. Points between the two thresholds only if they connect to a certain edge.

The fourth step is the real gain: it joins broken lines while suppressing isolated noise points, which a single threshold cannot do simultaneously.

Measuring with a reference#

The "same plane" restriction is decisive. Objects closer to the camera appear larger, and without rectification the measurement is wrong by the perspective factor. With a perpendicular shot and a reference next to the part, accuracies below one percent are attainable; with an angled shot and no rectification, ten percent errors are easy.