A Deformable Attention-Based Detection Transformer with Cross-Scale Feature Fusion for Industrial Coil Spring Inspection

Matteo Rossi; Pony Matt

cs.CVMar 13, 2026

A Deformable Attention-Based Detection Transformer with Cross-Scale Feature Fusion for Industrial Coil Spring Inspection

Authors: Matteo Rossi, Pony Matt

Organizations: Maharaja Agrasen University

Abstract

Automated visual inspection of locomotive coil springs presents significant challenges due to the morphological diversity of surface defects, substantial scale variations, and complex industrial backgrounds. This paper proposes MSD-DETR (Multi-Scale Deformable Detection Transformer), a novel detection framework that addresses these challenges through three key innovations: (1) a structural re-parameterization strategy that decouples training-time multi-branch topology from inference-time efficiency, enhancing feature extraction while maintaining real-time performance; (2) a deformable attention mechanism that enables content-adaptive spatial sampling, allowing dynamic focus on defect-relevant regions regardless of morphological irregularity; and (3) a cross-scale feature fusion architecture incorporating GSConv modules and VoVGSCSP blocks for effective multi-resolution information aggregation. Comprehensive experiments on a real-world locomotive coil spring dataset demonstrate that MSD-DETR achieves 92.4% [email protected] at 98 FPS, outperforming state-of-the-art detectors including YOLOv8 (+3.1% mAP) and the baseline RT-DETR (+2.8% mAP) while maintaining comparable inference speed, establishing a new benchmark for industrial coil spring quality inspection.

Figures & tables

Figure 1: Overall architecture of the proposed MSD-DETR framework. The backbone extracts hierarchical features enhanced through structural re-parameterization. The hybrid encoder processes multi-scale features using deformable attention. The bidirectional fusion neck aggregates cross-scale representations. The transformer decoder produces final detection outputs.

Category	Train	Val	Test	Total
Surface Crack	1,847	396	395	2,638
Corrosion	2,156	462	461	3,079
Decarburization	1,324	284	283	1,891
Surface Scratch	1,892	406	405	2,703
Pitting	1,778	381	377	2,536
Total	8,997	1,929	1,921	12,847

Table 1: Dataset statistics showing distribution of defect instances across categories and data splits.

Method	mAP	mAP	AP S	AP M	AP L	Params(M)	FLOPs(G)	FPS	Latency(ms)
Faster R-CNN	82.3	54.1	41.2	58.3	67.8	41.5	180.2	18	55.6
YOLOv5-L	86.7	58.3	45.8	62.1	71.5	46.5	109.1	82	12.2
YOLOv7	87.4	59.2	46.3	63.5	72.1	36.9	104.7	78	12.8
YOLOv8-L	88.2	60.1	47.5	64.2	73.0	43.7	165.2	78	12.8
DETR	79.8	51.2	38.6	55.4	64.2	41.0	86.0	25	40.0
Deformable DETR	85.4	57.8	44.1	61.8	70.3	40.0	78.0	42	23.8

Table 2: Performance comparison with state-of-the-art methods on the locomotive coil spring defect dataset. Best results are shown in bold , second-best are underlined .

Figure 2: Qualitative comparison of detection results. (a) Input images with ground truth annotations. (b) RT-DETR baseline predictions. (c) MSD-DETR predictions. Green boxes indicate correct detections, red boxes indicate false positives, and yellow boxes indicate missed detections.

Rep	DA	CSFF	[email protected]	[email protected]:0.95	FPS
			89.1	61.5	114
✓			90.3 (+1.2)	62.8 (+1.3)	112
	✓		90.8 (+1.7)	63.4 (+1.9)	102
		✓	90.5 (+1.4)	63.1 (+1.6)	108
✓	✓		91.6 (+2.5)	64.5 (+3.0)	100
✓		✓	91.2 (+2.1)	64.1 (+2.6)	104

Table 3: Ablation study results analyzing individual component contributions. “Rep” denotes structural re-parameterization, “DA” indicates deformable attention enhancement, and “CSFF” represents cross-scale feature fusion.

$K$	$L$	[email protected]	[email protected]:0.95	FPS
2	4	91.6	64.8	108
4	4	92.4	65.8	98
8	4	92.5	65.9	82
4	3	91.8	65.0	106
4	5	92.3	65.6	91

Table 4: Analysis of deformable attention configuration parameters.

3x3	1x1	Id	[email protected]	[email protected]:0.95
✓			89.1	61.5
✓	✓		89.7	62.1
✓		✓	89.5	61.9
✓	✓	✓	90.3	62.8

Table 5: Analysis of structural re-parameterization branch configurations. “3x3” denotes the primary convolution branch, “1x1” denotes the linear branch, and “Id” denotes the identity branch.

Defect Category	RT-DETR	MSD-DETR	Improvement
Surface Crack	87.2	91.8	+4.6
Corrosion	91.5	94.2	+2.7
Decarburization	85.3	89.6	+4.3
Surface Scratch	89.8	93.5	+3.7
Pitting	91.7	92.9	+1.2
Mean (mAP)	89.1	92.4	+3.3

Table 6: Per-category detection performance comparison ([email protected]).

Method	[email protected]	[email protected]:0.95
YOLOv8-L	62.4	38.7
RT-DETR-L	65.8	41.2
MSD-DETR (Ours)	68.3	43.9

Table 7: Cross-dataset generalization evaluation on NEU-DET steel surface defect dataset (zero-shot transfer).

Detection Transformer

Detection Transformer (DETR) uses transformer-based set prediction to detect objects end to end, without conventional non-maximum suppression.

See all 90 papers

ViCrop-Det: Spatial Attention Entropy Guided Cropping for Training-Free Small-Object Detection
Apr 29, 2026Hui Wang, Hongze Li, Wei Chen +1Small Object Detection Detection Transformer
School of Computer Science and Technology, Huazhong University of Science and Technology 2School of Software Engineering, Huazhong University of Science and Technology Wuhan 430074, China
Attention-Guided Autoencoder Fusion for Insulator Defect Detection Using UAV Transmission-Line Imaging
Jun 3, 2026Malak Allam, Khaled Shaban, Ali HamdiIndustrial Inspection Object Detection
Faculty of Computer Science MSA University Giza, Egypt · Department of Computer Science and Engineering Qatar University Doha, Qatar
SPDCN: Strip-based Deformable Convolutional Network for Steel Surface Defect Segmentation
Jul 23, 2026Zhongming Liu, Bingbing Jiang, Guangxin Wan +1Image Segmentation Industrial Inspection
School of Artificial Intelligence, Jiangxi Normal University, Nanchang, 330022, China · Jiangxi Provincial Key Laboratory of Intelligent Information Processing and Affective Computing, School of Artificial Intelligence, Jiangxi Normal University, Nanchang, 330022, China

Apr 29, 2026J/K move · Enter open · S save

ViCrop-Det: Spatial Attention Entropy Guided Cropping for Training-Free Small-Object Detection

Hui Wang, Hongze Li, Wei Chen, Xiaojin Zhang

School of Computer Science and Technology, Huazhong University of Science and Technology 2School of Software Engineering, Huazhong University of Science and Technology Wuhan 430074, China

Small Object Detection Detection Transformer Aerial Object Detection Attention Entropy

Open paper AskPDF arXiv

Transformer-based architectures have established a dominant paradigm in global semantic perception; however, they remain fundamentally constrained by the profound spatial heterogeneity inherent in natural images. Specifically, the imposition of a uniform global receptive field across regions of varying information density inevitably leads to local feature degradation, particularly in dense conflict zones populated by microscopic targets. To address this mechanistic limitation, we propose ViCrop-Det, a training-free inference framework that introduces adaptive spatial trust region shrinkage. Inspired by the use of attention entropy in anomaly segmentation, ViCrop-Det leverages the detection decoder's cross-attention distribution as an endogenous probe. By utilizing Spatial Attention Entropy (SAE) to heuristically evaluate local spatial ambiguity, the framework executes dynamic spatial routing, allocating a fixed computational budget exclusively to regions exhibiting both high target saliency and high cognitive uncertainty. By shrinking the spatial trust region and injecting high-frequency localized observations, ViCrop-Det actively resolves spatial ambiguity and recovers fine-grained features without requiring architectural modifications. Extensive evaluations on VisDrone and DOTA-v1.5 demonstrate that ViCrop-Det yields competitive performance enhancements, consistently adding +1-3 mAP@50 to RT-DETR-R50 and Deformable DETR with a marginal 20-23% latency overhead. On MS COCO,

AP_{S}

improves while

AP_{M}/AP_{L}

remains stable, indicating precise fine-scale refinement without compromising the global spatial prior. Under compute-matched settings, our adaptive routing strategy comprehensively surpasses uniform slicing baselines, achieving a highly optimized accuracy-speed trade-off.

Abstract

Detection Transformer

Explore similar work

ViCrop-Det: Spatial Attention Entropy Guided Cropping for Training-Free Small-Object Detection

Attention-Guided Autoencoder Fusion for Insulator Defect Detection Using UAV Transmission-Line Imaging

SPDCN: Strip-based Deformable Convolutional Network for Steel Surface Defect Segmentation