Principle:Kornia Kornia Deep Edge Detection

Knowledge Sources	Dense Extreme Inception Network: Towards a Robust CNN Model for Edge Detection Kornia
Domains	Vision Deep_Learning Edge_Detection
Last Updated	2026-02-09 15:00 GMT

Overview

Technique of using deep neural networks trained on edge annotations to produce high-quality edge maps that capture semantic boundaries beyond simple intensity gradients.

Description

Deep edge detection uses convolutional neural networks trained on human-annotated edge datasets (like BIPED) to detect semantically meaningful boundaries. Unlike classical methods (Canny, Sobel) that rely on intensity gradients, deep models learn to distinguish true object boundaries from texture edges and noise.

DexiNed (Dense Extreme Inception Network) uses a multi-scale architecture with dense blocks, skip connections, and multi-level feature fusion to produce accurate edge maps at the original image resolution. The network generates edge predictions at 6 different scales and fuses them through a learned combination layer.

Usage

Use when classical edge detection produces noisy results or misses semantic boundaries. DexiNed with pretrained weights works out-of-the-box for general edge detection without fine-tuning.

Theoretical Basis

The DexiNed architecture consists of 6 dense blocks with skip connections operating at multiple scales:

Architecture overview:
  Input (B, 3, H, W)
    -> Dense Block 1 (full resolution)
    -> Dense Block 2 (1/2 resolution)
    -> Dense Block 3 (1/4 resolution)
    -> Dense Block 4 (1/8 resolution)
    -> Dense Block 5 (1/16 resolution)
    -> Dense Block 6 (1/32 resolution)

Each block produces a side output:
  side_i = upsample_to_original(block_i_output)

Final fusion combines all scale predictions through a learned 1x1 convolution:

Failed to parse (syntax error): {\displaystyle \text{edge\_map} = \text{Conv}_{1 \times 1}(\text{cat}(\text{up}_1, \text{up}_2, \text{up}_3, \text{up}_4, \text{up}_5, \text{up}_6))}

Training loss uses binary cross-entropy with human-annotated edge labels:

${ℒ}=-\frac{1}{N}\sum _{i=1}^N[y_i\log (\sigma (p_i))+(1-y_i)\log (1-\sigma (p_i))]$

where $\sigma$ is the sigmoid function, $p_i$ is the predicted logit, and $y_i$ is the ground-truth edge label.

Related Pages

Implemented By

• Implementation:Kornia_Kornia_DexiNed_Detector