PY-L5-30 · OpenCV Basics — Load, Display, Resize, Filter

Learning Goals

3 min

Read and write images with cv2.imread / imwrite.
Remember OpenCV uses BGR, not RGB.
Resize, crop, convert to grayscale.
Apply blur and Canny edge detection.

Warm-Up · BGR, the Famous Gotcha

5 min

pip install opencv-python

import cv2
img = cv2.imread("cat.jpg")     # loads as a NumPy array (Lesson 5!)
print(img.shape)                # (H, W, 3)
print(img[0, 0])                # [B G R]  ← Blue, Green, Red order!

OpenCV stores colours as BGR, not RGB — a historical quirk. If your displayed image looks blue-tinted, you forgot to convert. To show with matplotlib (which expects RGB): cv2.cvtColor(img, cv2.COLOR_BGR2RGB).

Today's big idea

An OpenCV image IS a NumPy array — everything from Lesson 5 applies. OpenCV just adds fast, battle-tested operations (resize, blur, edges, detection) on top. Watch the BGR/RGB order.

New Concept · The Core Operations

14 min

Read, write, show

import cv2, matplotlib.pyplot as plt

img = cv2.imread("cat.jpg")
cv2.imwrite("copy.png", img)              # save

# show via matplotlib (convert BGR→RGB first)
plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB)); plt.axis("off")

Resize & crop

small = cv2.resize(img, (200, 200))          # exact size
half  = cv2.resize(img, None, fx=0.5, fy=0.5) # by factor
crop  = img[50:250, 100:400]                  # NumPy slice

Grayscale

gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)   # shape (H, W)

Blur (smoothing)

blur = cv2.GaussianBlur(img, (7, 7), 0)   # kernel must be odd

Blurring averages nearby pixels — used to reduce noise before edge detection, and for privacy (blur faces/plates).

Edges with Canny

edges = cv2.Canny(gray, threshold1=100, threshold2=200)
# white pixels where the image changes sharply

Draw on images

cv2.rectangle(img, (50, 50), (200, 200), (0, 255, 0), 2)   # green box
cv2.putText(img, "cat", (50, 40), cv2.FONT_HERSHEY_SIMPLEX,
            1, (0, 255, 0), 2)

These draw functions are how detection results get annotated (Lesson 31's face boxes).

Worked Example · A Mini Vision Pipeline

12 min

# vision_basics.py — load → gray → blur → edges, side by side
import cv2, matplotlib.pyplot as plt

img  = cv2.imread("street.jpg")
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
blur = cv2.GaussianBlur(gray, (5, 5), 0)
edges = cv2.Canny(blur, 80, 160)

panels = {
    "original": cv2.cvtColor(img, cv2.COLOR_BGR2RGB),
    "gray":     gray,
    "blurred":  blur,
    "edges":    edges,
}
fig, ax = plt.subplots(1, 4, figsize=(16, 4))
for a, (name, im) in zip(ax, panels.items()):
    a.imshow(im, cmap="gray" if im.ndim == 2 else None)
    a.set_title(name); a.axis("off")
fig.tight_layout(); fig.savefig("pipeline.png", dpi=130)

Read the diff

The canonical pre-processing chain: colour → gray (simplifies) → blur (denoise) → Canny (find edges). Blurring before Canny is important — raw edges are noisy. This exact pipeline feeds lane detection, document scanning, and shape finding. Each step is one OpenCV call on a NumPy array.

Try It Yourself

13 min

01 🟢 Resize & save

Load a photo, resize it to 300px wide (keep aspect ratio), and save it. Print original and new shapes.

Hint

h, w = img.shape[:2]
new_w = 300
new_h = int(h * new_w / w)
out = cv2.resize(img, (new_w, new_h))
cv2.imwrite("resized.jpg", out)

02 🟡 Edge slider

Try three Canny threshold pairs (low/high) and compare the edge maps. How do thresholds change what counts as an edge?

03 🔴 Blur a region

Blur just a rectangular region of an image (e.g., to censor it) and paste it back. Combine slicing + GaussianBlur.

Hint

x1, y1, x2, y2 = 100, 80, 260, 220
region = img[y1:y2, x1:x2]
img[y1:y2, x1:x2] = cv2.GaussianBlur(region, (35, 35), 0)

Mini-Challenge · Document Scanner Look

8 min

Give a photo the "scanned document" look: grayscale, then adaptive threshold to make it crisp black-and-white. Compare to a plain global threshold.

Show one possible solution

import cv2
gray = cv2.cvtColor(cv2.imread("note.jpg"), cv2.COLOR_BGR2GRAY)
scan = cv2.adaptiveThreshold(gray, 255,
        cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 11, 2)
cv2.imwrite("scanned.png", scan)

Adaptive thresholding adjusts per-region — it handles uneven lighting that a single global threshold can't. That's why scanner apps look clean across the whole page.

Recap

3 min

OpenCV images are NumPy arrays in BGR order. Core ops: imread/imwrite, resize, cvtColor (gray/RGB), GaussianBlur, Canny edges, and draw functions. The classic chain is gray → blur → edges. Adaptive thresholding beats global on uneven lighting. Next: detect faces with one library call.

Vocabulary Card

BGR: OpenCV's colour order (Blue, Green, Red) — convert to RGB before matplotlib.
cvtColor: Convert between colour spaces (BGR↔RGB, BGR↔GRAY).
GaussianBlur: Smoothing filter; reduces noise (odd kernel size).
Canny: Edge-detection algorithm with low/high thresholds.

Homework

4 min

Build cv_toolkit.py with functions: resize_to_width, to_gray, blur, edges, and censor_region. Apply all five to a photo and save a labelled grid PNG. Add a one-line docstring per function.

import cv2, matplotlib.pyplot as plt img = cv2.imread("cat.jpg") cv2.imwrite("copy.png", img) # save # show via matplotlib (convert BGR→RGB first) plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB)); plt.axis("off")

# vision_basics.py — load → gray → blur → edges, side by side import cv2, matplotlib.pyplot as plt img = cv2.imread("street.jpg") gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) blur = cv2.GaussianBlur(gray, (5, 5), 0) edges = cv2.Canny(blur, 80, 160) panels = { "original": cv2.cvtColor(img, cv2.COLOR_BGR2RGB), "gray": gray, "blurred": blur, "edges": edges, } fig, ax = plt.subplots(1, 4, figsize=(16, 4)) for a, (name, im) in zip(ax, panels.items()): a.imshow(im, cmap="gray" if im.ndim == 2 else None) a.set_title(name); a.axis("off") fig.tight_layout(); fig.savefig("pipeline.png", dpi=130)

import cv2 gray = cv2.cvtColor(cv2.imread("note.jpg"), cv2.COLOR_BGR2GRAY) scan = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 11, 2) cv2.imwrite("scanned.png", scan)