Sovereign 8s.is mirror of standardgalactic/C-TDED
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-03-22 22:06:50 +08:00
edge_losses.py Add files via upload 2025-10-26 20:16:38 +08:00
model.py Add files via upload 2025-10-26 20:16:38 +08:00
model_calculate.py Remove outdated comment from model_calculate.py 2026-03-22 22:05:27 +08:00
README.md Add algorithm details to README 2026-03-09 17:30:29 +08:00
readme.md Add files via upload 2025-10-26 20:16:38 +08:00
requirements.txt Add files via upload 2025-10-26 20:16:38 +08:00
training.py Add files via upload 2025-10-26 20:16:38 +08:00

C-TDED

High-Resolution Edge Detection Model Based on Dual-Branch Attention Fusion

Python PyTorch License

Overview | 概述

C-TDED is a high-performance dual-branch edge detection model that effectively integrates CNN-based branch for fine details with Transformer-based branch for global context.

C-TDED 是一个高性能的双分支边缘检测模型,有效地整合了基于CNN的细节分支和基于Transformer的全局上下文分支。

Model Architecture | 模型架构

  • Fine Semantic Edge Branch (FSEB): Captures high-resolution edge features using modified EfficientNet-B0
  • Global Context Branch (GCB): Models global semantic information using lightweight Transformer
  • Memory-Efficient Fusion Module (MEFM): Integrates features from both branches

  • 细粒度语义边缘分支 (FSEB):使用改进的EfficientNet-B0捕获高分辨率边缘特征
  • 全局上下文分支 (GCB):使用轻量级Transformer建模全局语义信息
  • 内存高效融合模块 (MEFM):整合两个分支的特征

Installation | 安装

# Clone the repository | 克隆仓库
git clone https://github.com/Ilk-cq/C-TDED
cd C-TDED

# Install dependencies | 安装依赖
pip install -r requirements.txt

Datasets | 数据集

This model has been evaluated on three benchmark datasets:

本模型在三个基准数据集上进行了评估:

BSDS500

  • Download | 下载: Berkeley Segmentation Dataset
  • Description | 描述: 500 natural images with human-annotated boundaries
  • 结构: 500张带有人工标注边界的自然图像

Multicue

  • Download | 下载: Multicue Dataset
  • Description | 描述: Dataset for boundary detection with multiple visual cues
  • 结构: 用于边界检测的多视觉线索数据集

NYUDv2

  • Download | 下载: NYU Depth Dataset V2
  • Description | 描述: Indoor scene dataset with RGB-D images
  • 结构: 包含RGB-D图像的室内场景数据集

Usage | 使用方法

Training | 训练

from training import train_edge_detection_model

# Train on BSDS500 dataset | 在BSDS500数据集上训练
checkpoint_dir = train_edge_detection_model(
    dataset='bsds500',
    data_root='path/to/bsds500',
    epochs=100,
    batch_size=8,
    learning_rate=1e-4,
    img_size=224
)

Inference | 推理

import torch
from model import DualBranchEdgeModel, inference_with_sigmoid

# Load model | 加载模型
model = DualBranchEdgeModel(img_size=224)
model.load_state_dict(torch.load('checkpoint.pth')['model_state_dict'])
model.eval()

# Inference | 推理
with torch.no_grad():
    outputs = inference_with_sigmoid(model, input_image)
    edge_map = outputs['final_edges']

Performance | 性能

Dataset ODS OIS
BSDS500 0.847 0.861
Multicue (Edge) 0.899 0.907
Multicue (Boundary) 0.842 0.850
NYUDv2 0.761 0.776

Key Features | 主要特点

  • High-Resolution Feature Extraction | 高分辨率特征提取: Modified backbone with reduced initial downsampling
  • Efficient Local Attention | 高效局部注意力: Window-based attention mechanism (7×7)
  • Structure-Aware Loss | 结构感知损失: Comprehensive loss function including gradient, continuity, and directional consistency
  • Multi-Scale Supervision | 多尺度监督: Side outputs at different scales for deep supervision

  • 高分辨率特征提取:改进的骨干网络,减少初始下采样
  • 高效局部注意力:基于窗口的注意力机制(7×7)
  • 结构感知损失:包括梯度、连续性和方向一致性的综合损失函数
  • 多尺度监督:不同尺度的侧输出用于深度监督

File Structure | 文件结构

C-TDED/
├── model.py              # Model architecture | 模型架构
├── dataset.py            # Dataset loaders | 数据集加载器
├── edge_losses.py        # Loss functions | 损失函数
├── training.py           # Training script | 训练脚本
├── model_calculate.py    #Calculating model parameters and others
├── requirements.txt      # Dependencies | 依赖项
└── README.md            # This file | 本文件

Requirements | 环境要求

  • Python >= 3.8
  • PyTorch >= 1.10.0
  • CUDA >= 11.0 (for GPU support | GPU支持)
  • See requirements.txt for complete list | 完整列表见 requirements.txt

##其他算法详情 HED - Holistically-nested Edge Detection Full name: Holistically-nested Edge Detection (HED) GitHub: https://github.com/s9xie/hed Paper: ICCV 2015

BDCN - Bi-Directional Cascade Network Full name: Bi-Directional Cascade Network for Perceptual Edge Detection (BDCN) GitHub: https://github.com/pkuCactus/BDCN Paper: CVPR 2019

EDTER - Edge Detection with TRansformer Full name: Edge Detection with Transformer (EDTER) GitHub: https://github.com/MengyangPu/EDTER Paper: CVPR 2022

DiffusionEdge - Diffusion Probabilistic Model for Crisp Edge Detection Full name: DiffusionEdge: Diffusion Probabilistic Model for Crisp Edge Detection GitHub: https://github.com/GuHuangAI/DiffusionEdge Paper: AAAI 2024

License | 许可证

This project is licensed under the MIT License.

本项目采用 MIT 许可证。