Edge Detection

Preface

This tutorial applies to firmware version: CanMV_K230_YAHBOOM_micropython_v1.8-13. You can download it from the Jianan website https://download.kendryte.com/developer/releases/canmv_k230_micropython/daily_build/

Example Results

Run the example code for this section 11.camera_edge_detect.py.

This section first converts the live color image to grayscale, then demonstrates Canny, Sobel, Scharr, and Laplacian edge or gradient operators. The output is displayed as a single-channel grayscale image.

Principle Explanation

First and Second Order Derivatives

Image edges correspond to locations where grayscale changes rapidly. Sobel and Scharr approximate first-order derivatives, detecting gradients in the x or y direction respectively. Laplacian computes second-order derivatives, which is more sensitive to details and noise.

Canny is a multi-stage algorithm that typically includes smoothing, gradient computation, non-maximum suppression, and double threshold linking, outputting relatively thin and continuous binary edges.

Code Overview

Initialize Camera and Display Module

W, H = 640, 480
sensor = Sensor(width=1280, height=960, fps=90)
sensor.reset()
sensor.set_framesize(width=W, height=H)
sensor.set_pixformat(Sensor.RGB888)
Display.init(Display.ST7701, width=640, height=480, to_ide=True)
sensor.run()
time.sleep(0.5)

Set Detection Mode

xxxxxxxxxx
mode = 0

Convert to Grayscale

xxxxxxxxxx
img = sensor.snapshot()
img_np = img.to_numpy_ref()
gray = cv2.cvtColor(img_np, cv2.COLOR_BGR2GRAY)

Edge detection usually does not need color information. Grayscale reduces data volume and unifies channel processing.

Execute Edge or Gradient Computation

xxxxxxxxxx
if mode == 0:
    out = cv2.Canny(gray, 80, 200)
elif mode == 1:
    out = cv2.Sobel(gray, cv2.CV_8U, 1, 0)
elif mode == 2:
    out = cv2.Scharr(gray, cv2.CV_8U, 1, 0)
else:
    out = cv2.Laplacian(gray, cv2.CV_8U)

The current example directly outputs CV_8U. Negative gradients cannot be fully preserved in unsigned 8-bit output, so results are suitable for visual demonstration but not for direct precise gradient numerical analysis.

Package Single-Channel Result and Display

xxxxxxxxxx
img_out = image.Image(
    W, H, image.GRAYSCALE,
    alloc=image.ALLOC_REF,
    data=out
)
Display.show_image(img_out)

Resource Release

xxxxxxxxxx
finally:
    sensor.stop()
    Display.deinit()
    os.exitpoint(os.EXITPOINT_ENABLE_SLEEP)
    time.sleep_ms(100)

Parameter Adjustment Notes

• Canny Double Threshold: The high threshold determines strong edges, and the low threshold is used to connect weak edges adjacent to strong edges. The low threshold is usually less than the high threshold.
• Pre-denoising: The current script does not have Gaussian preprocessing. When there are many noise edges, you can first add 3×3 or 5×5 Gaussian blur.
• Gradient Direction: Sobel/Scharr with dx=1, dy=0 computes changes in the x direction, highlighting edges close to vertical. Changing to dx=0, dy=1 allows observation of the other direction.
• Output Depth: When you need to preserve positive and negative gradients, use firmware-supported signed or floating-point depths, then take the absolute value and normalize for display.
• Laplacian Noise: Second-order derivatives amplify high-frequency noise, usually relying more on pre-smoothing.
• Performance: You can downsample when only doing detection. When you need to draw edge coordinates back onto the original image, you must synchronously scale the coordinates.