Sovereign 8s.is mirror of standardgalactic/DM-YOLO
  • Python 99.8%
  • Shell 0.2%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-08-23 18:52:52 +08:00
ultralytics Initial commit 2025-09-10 17:25:29 +08:00
.gitattributes Initial commit 2025-09-10 17:25:29 +08:00
FPS.py Initial commit 2025-09-10 17:25:29 +08:00
LICENSE Create LICENSE 2025-09-10 17:34:14 +08:00
README.md Update README.md 2026-08-23 18:52:52 +08:00
test.py Initial commit 2025-09-10 17:25:29 +08:00
track.py Initial commit 2025-09-10 17:25:29 +08:00
train.py Initial commit 2025-09-10 17:25:29 +08:00
yolov8n.pt Initial commit 2025-09-10 17:25:29 +08:00

1. Code Dependencies and Requirements

1.1 Experimental Platform and Environment

  • CPU: Intel Xeon Platinum 8369B
  • GPU: NVIDIA A10
  • Python: 3.9.0
  • Pytorch: 2.1.0
  • CUDA: 11.8

1.2 Training Hyperparameters

  • Decay strategy: cosine
  • Optimizer: SGD
  • Learning rate: 0.01
  • Weight decay: 0.0005
  • Momentum: 0.937
  • Close mosaic: 10
  • Total epochs: 500
  • Batch size: 64
  • Patience: 100

1.3 Requirements

  • faster-coco-eval>=1.6.5
  • PyYAML
  • tensorboard
  • scipy
  • calflops
  • thop
  • transformers
  • pytorch_wavelets==1.3.0
  • timm==1.0.7
  • grad-cam==1.5.4
  • tidecv
  • einops

2. File Description

  • Model configuration file: ultralytics/yaml/v8-n-DRF-MFF.yaml
  • train.py: Script for training the model
  • track.py: Inference script
  • test.py: Testing script
  • FPS.py: Script for calculating model inference time and FPS

3. Key Modules of the Algorithm

3.1 Dynamic Receptive Field (DRF)

Code: ultralytics/nn/modules/DRF.py

Description: Dynamic Receptive Field (DRF) module, which enables the network to dynamically adjust its receptive field according to defect scales, thereby enabling effective extraction of multi-scale feature representations.

3.2 Multi-scale Feature Fusion (MFF)

Code: ultralytics/nn/modules/MFF.py

Description: Multi-scale Feature Fusion (MFF) module combines shallow and deep contextual information, minimizing information loss and enhancing feature representation.


4. Related Details

This paper conducts experiments using three popular public datasets: NEU-DET, GC10-DET, and APDDD.

All datasets are divided into training, validation, and test sets in an 8:1:1 ratio.

The relevant datasets can be obtained from the Baidu Netdisk link:
Baidu Netdisk Link
Extraction code: u5rs

If you have any questions or needs, please contact me:
Email: xs1113@stu.ahjzu.edu.cn


5. Citation Format

Article title: Steel Surface Defect Detection Based on Dynamic Receptive Field and Multi-Scale Features Fusion

Submitted journal: Scientific Reports