SLAT-YOLO

English | 简体中文

SLAT-YOLO is a YOLO11-based detector for track slab cracks. It combines attention-enhanced backbone features, token refinement, lightweight feature fusion and context-aware detection. This repository provides model code, YAML configurations, training scripts, experimental curves and baseline/ablation checkpoints, built on Ultralytics 8.4.143.

Associated paper: SLAT-YOLO: Shape-aware attention and spatially enhanced token refinement for robust track slab crack detection.

Modules

Module	Role	Implementation
C3k2-CA	Adds coordinate attention to backbone features to retain directional and positional information.	C3k2, CA
C2-STR	Refines feature tokens through token statistics attention, Dynamic Tanh, Mona adaptation and gated feed-forward processing.	transformer.py
VoV-GSCSP / GSConv	Combines lightweight convolution and cross-stage feature fusion in the neck.	block.py
MPCR-Head	Recalibrates box and classification features using context from multiple patch scales.	head.py
Wise-ShapeIoU	Combines shape-aware box regression with adaptive loss weighting.	slat_loss.py

Model and ablation configurations are in my_project/yaml. The default entry is slat-yolo.yaml.

Installation and training

Use Python 3.10+ and a PyTorch installation appropriate for your CPU/CUDA environment.

git clone https://github.com/cjdfree/SLAT-YOLO.git
cd SLAT-YOLO
pip install -e .

The research dataset is not distributed. Supply your own YOLO dataset YAML using --data.

python -m my_project.run train --data /path/to/your/data.yaml --device 0 --imgsz 1024

# YOLO11n baseline
python -m my_project.run train --model my_project/yaml/yolo11n.yaml --data /path/to/your/data.yaml --device 0 --name baseline

The training script sets the following project defaults; unspecified options use Ultralytics defaults. The image input size is 1024. Outputs are saved under my_project/outputs/.

Parameter	Default
imgsz	1024
epochs / batch	1200 / 16
optimizer / lr0 / lrf	auto / 0.01 / 0.01
momentum / weight_decay	0.937 / 0.0005
pretrained / amp	False / False
seed / deterministic	0 / True
patience / workers	0 / 4

Experimental results

Values below are percentages from the epoch with the highest recorded validation AP50 in each training CSV; precision, recall and AP50–95 come from the same epoch. These are archived validation results, not independent test scores. The historical split contains augmented images derived from shared originals across training and validation.

Ablation study

Rows denote additions to YOLO11n; SLAT-YOLO denotes the full configuration.

Configuration	Precision (%)	Recall (%)	AP50 (%)	AP50–95 (%)
C2-STR	94.31	88.09	93.78	68.48
C3k2-CA	95.06	83.62	92.81	72.93
C3k2-CA-C2-STR	97.17	85.11	94.71	69.50
C3k2-CA-C2-STR-VoV-GSCSP	95.03	89.58	95.75	78.69
C3k2-CA-C2-STR-VoV-GSCSP-MPCR-Head	96.79	89.57	96.10	78.12
SLAT-YOLO	97.52	91.98	96.62	79.62
C3k2-CA-MPCR-Head	95.69	87.45	94.47	73.95
C3k2-CA-VoV-GSCSP	95.21	89.79	95.38	71.43
C3k2-CA-VoV-GSCSP-MPCR-Head	94.25	91.70	96.87	79.25
C3k2-CA-VoV-GSCSP-MPCR-Head-CIoU	97.50	91.28	95.84	77.74
C3k2-CA-VoV-GSCSP-MPCR-Head-GIoU	95.45	91.28	95.93	80.36
C3k2-CA-VoV-GSCSP-MPCR-Head-SIoU	95.44	88.99	95.52	73.93
C3k2-CA-VoV-GSCSP-MPCR-Head-Wise-ShapeIoU	94.35	92.55	96.14	78.36
MPCR-Head	95.83	81.70	91.87	69.48
VoV-GSCSP	89.74	84.68	91.79	67.02
VoV-GSCSP-MPCR-Head	94.84	87.66	94.15	73.66
Wise-ShapeIoU	92.77	86.81	93.42	72.50

Loss comparison

Configuration	Precision (%)	Recall (%)	AP50 (%)	AP50–95 (%)
CIoU	94.04	82.13	90.26	69.53
DIoU	94.21	72.69	85.50	61.60
EIoU	94.59	81.78	89.43	64.98
GIoU	93.87	79.36	86.71	66.72
Inner-CIoU	93.73	85.11	90.46	69.74
Inner-PIoU	95.57	82.13	90.74	69.52
Inner-ShapeIoU	91.23	75.96	85.03	62.70
PIoU	94.42	84.89	92.23	69.51
ShapeIoU	92.39	84.47	90.76	64.66
SIoU	95.04	81.55	90.52	70.33
Wise-CIoU	94.07	85.32	92.00	69.63
Wise-Inner-MPDIoU-v3	93.81	85.11	91.52	70.90
Wise-ShapeIoU	92.77	86.81	93.42	72.50
Wise-WIoU	93.77	83.83	92.16	70.29

Attention comparison

Configuration	Precision (%)	Recall (%)	AP50 (%)	AP50–95 (%)
C3k2-CA	95.06	83.62	92.81	72.93
CBAM	95.03	81.39	91.02	67.10
EMA	93.54	78.51	88.97	61.67
SEAttention	92.90	83.47	91.50	71.59

Detector comparison

Configuration	Precision (%)	Recall (%)	AP50 (%)	AP50–95 (%)
RT-DETR-L	95.05	75.53	82.93	54.83
YOLO11n	96.98	82.12	91.30	70.02
YOLO12n	98.56	80.00	89.94	68.44
YOLO13n	94.96	83.40	90.12	71.35
YOLOv10n	90.38	79.95	89.51	59.56
YOLOv5n	93.74	73.40	82.04	61.88
YOLOv8n	95.31	82.12	89.75	69.46
YOLOv9t	85.99	73.15	79.91	51.18

Training durations differ: YOLOv5n has 2500 recorded epochs, RT-DETR-L has 515, and the other comparison runs have 1200. All tables link to their training logs and curves. Full summaries are available in ablation_summary.csv and comparison_summary.csv. Image previews and annotations are excluded.

Checkpoints and prediction

The downloader provides 17 baseline/ablation inference checkpoints. The final full-model experiment has training logs but no released checkpoint; the example below uses the architecture ablation without the final Wise-ShapeIoU stage.

python -m my_project.download_assets checkpoints
python -m my_project.run predict --model my_project/checkpoints/yolo11n-C3k2-CA-C2-STR-VoV-GSCSP-MPCR-Head.pt --source /path/to/images --device 0 --imgsz 1024

License and acknowledgments

This repository uses AGPL-3.0. Citation metadata is provided in CITATION.cff. Building blocks were adapted from the authors' experiment archives and the following projects:

Ultralytics, Coordinate Attention, ToST, Mona, GSConv, SEAM / MultiSEAM.

Original third-party notices are retained in docs/upstream/licenses. SLAT-YOLO is a research project and is not an official Ultralytics product.