This repository provides the core implementation of the key modules proposed in our paper:

ESA-Det: Enhancing Remote Sensing Object Detection via Edge-Prior Infused Attention and Spatial Adaptive Fusion

To promote reproducibility and further research, we release lightweight and modular implementations of the two main contributions:

• Edge-Prior Infused Attention (EPIA)
• Spatial Adaptive Fusion (SAF) Block

⚠️ Note: The full training pipeline is currently being organized and will be released in a subsequent update. This repository focuses on clean, minimal, and reusable core modules.

File: epia.py Core Module: C3k2_EPIA

EPIA integrates classical edge operators into CNN feature learning to explicitly enhance boundary-aware representations.

• Multi-operator edge extraction:

  • Sobel
  • Prewitt
  • Scharr
  • Robert

• Depthwise convolution implementation (efficient & parameter-free)

• Residual fusion with learnable convolution branch

Remote sensing objects often exhibit:

• weak boundaries
• low contrast
• cluttered backgrounds

EPIA introduces explicit structural priors to address these issues.

File: saf.py Core Module: A2C2f_SAF

SAF follows a three-stage pipeline:

1. AFC (Calibration) Adaptive nonlinear feature normalization

2. Fusion Gate Frequency-aware spatial fusion

3. FEA (Enhancement) Multi-scale structural refinement

• Lightweight design

• Spatially adaptive feature fusion

• Strong performance on:

  • small objects
  • dense scenes
  • low-contrast targets

git clone https://github.com/yourname/ESA-Det.git
cd ESA-Det
pip install -r requirements.txt

from epia import C3k2_EPIA
from saf import A2C2f_SAF

# Replace backbone block
backbone_block = C3k2_EPIA(c1=256, c2=256, n=2)

# Replace neck/head block
fusion_block = A2C2f_SAF(c1=512, c2=512, n=2)

This project is released under the AGPL-3.0 License.