Implementation:Kornia Kornia Edge Map Processing

Knowledge Sources	Kornia
Domains	Vision Image_Processing
Last Updated	2026-02-09 15:00 GMT

Overview

Pattern for post-processing edge detection outputs using PyTorch operations and Kornia's EdgeDetector visualization.

Description

This documents the common patterns for edge map post-processing. Thresholding uses torch.threshold or comparison operators on gradient magnitudes. Sigmoid normalization uses torch.sigmoid on raw logits from deep models. The EdgeDetector.visualize method converts edge maps to RGB images or PIL Images for display. These are user-applied patterns rather than a single API.

Usage

Apply after edge detection to prepare outputs for visualization or downstream processing.

Code Reference

Source Location

Repository: kornia
File: kornia/filters/canny.py (hysteresis L132-159), kornia/contrib/edge_detection.py (visualize L118-141)

Signature

Common interface patterns:

# Binary thresholding
binary_edges = (magnitude > threshold).float()

# Sigmoid normalization
normalized = torch.sigmoid(raw_logits)

# EdgeDetector visualization
rgb_edges = EdgeDetector.visualize(images, edge_maps, output_type="torch")

Import

from kornia.contrib import EdgeDetector  # for visualize method
import torch  # for torch.threshold, torch.sigmoid

I/O Contract

Inputs

Name	Type	Required	Description
raw_edge_map	torch.Tensor	Yes	Edge detection output of shape (B, 1, H, W) or (B, C, H, W)
threshold	float	No	Threshold value for binary edge conversion
images	torch.Tensor	No	Original images for visualization overlay

Outputs

Name	Type	Description
binary_edges	torch.Tensor	Binary edge map of shape (B, 1, H, W) with values 0 or 1
normalized_map	torch.Tensor	Sigmoid-normalized edge map in [0, 1] range
rgb_visualization	PIL.Image	RGB visualization of edges for display

Usage Examples

Thresholding Sobel Output

import torch
from kornia.filters import Sobel, Canny

# Thresholding Sobel output
sobel = Sobel()
img = torch.rand(1, 1, 256, 256)
magnitude = sobel(img)
binary_edges = (magnitude > 0.1).float()

# Canny already includes thresholding
canny = Canny(low_threshold=0.1, high_threshold=0.3)
mag, edges = canny(img)
# edges is already binary from hysteresis

Sigmoid Normalization of DexiNed Output

import torch
from kornia.filters import DexiNed

# Get raw logits from DexiNed
model = DexiNed(pretrained=True).eval()
img = torch.rand(1, 3, 352, 352)
with torch.no_grad():
    logits = model(img)

# Normalize to [0, 1] range
edge_map = torch.sigmoid(logits)

# Convert to binary with threshold
binary = (edge_map > 0.5).float()

EdgeDetector Visualization

from kornia.contrib import EdgeDetectorBuilder

detector = EdgeDetectorBuilder.build(pretrained=True, image_size=352)
img = torch.rand(1, 3, 480, 640)
edges = detector(img)

# Get PIL images for display
pil_images = detector.visualize(img, edges, output_type="pil")

# Get torch tensors for further processing
torch_vis = detector.visualize(img, edges, output_type="torch")

Related Pages

Implements Principle

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment