CF-YOLO: Context-Aware Feature Refinement for Camouflaged Industrial Micro-Defect Detection

Xinda Yu; Kunxin Zheng; Chunan Yu; Qingbo Song; Hao Xiao; Ying Zang; Jie Liu

cs.CVAug 28, 2026

CF-YOLO: Context-Aware Feature Refinement for Camouflaged Industrial Micro-Defect Detection

Authors: Xinda Yu, Kunxin Zheng, Chunan Yu, Qingbo Song, Hao Xiao, Ying Zang, Jie Liu

Organizations: School of Information Engineering, Huzhou University, Huzhou, 313000, China. · School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, 210000, China.

Abstract

Automated detection of surface micro-defects on industrial components, such as copper tubes, is critically important for quality assurance but remains challenging due to the minute scale of anomalies and their visual camouflage against complex backgrounds. These factors lead to weak feature representations and high rates of false positives and missed detections. To address these issues, we propose a novel real-time detection framework designed for efficient context perception and feature refinement. Our method integrates a Context-Perception Aggregation Module (CPAM), which synergises large-kernel perception for macro-texture context and small-kernel aggregation for sharp boundary delineation, effectively breaking the background camouflage. Furthermore, a Feature Additive Refinement Module (FARM) employs a linear-complexity additive token mixer to globally verify and refine the representation of fine-grained anomalies, suppressing noise-induced errors. To support research in this domain, we introduce the Copper Tube Defect Dataset (CTDD), a manually annotated benchmark containing 1,847 images and 4,898 boundingbox defect instances from copper-tube inspection scenarios. Extensive experiments demonstrate that our detector achieves strong and consistent performance on CTDD, outperforming representative baseline detectors, including YOLOv11, by 2.2% in mAP@50 and 3.9% in Precision while maintaining real-time inference speed. This work provides a robust and efficient solution for high-precision industrial inspection, bridging the gap between contextual understanding and detailed feature analysis. Our code and model are available at: https://github.com/Yu-Xinda/CFYOLO-Context-Aware-Feature-Refinement-for-Camouflaged-Industrial-Micro-Defect-Detection

Figures & tables

Figure 1: Visualization of challenging defects with severe background camouflage. The anomalies are nearly indistinguishable due to highly similar textures with the background. Their locations are indicated by red bounding boxes.

Figure 2: Schematic illustration of CF-YOLO for industrial copper tube defect detection, integrating CPAM for context perception and FARM for semantic refinement.

Figure 3: The architecture of the Context-Perception Aggregation Module (CPAM), illustrating the decoupling of Large-Kernel Perception (LKP) and Small-Kernel Aggregation (SKA).

Figure 4: The architecture of the Feature Additive Refinement Module (FARM). Adopting the Convolutional Additive Token Mixer (CATM) paradigm, it performs global semantic refinement via dual-domain (spatial and channel) interactions and linear-complexity additive similarity.

Figure 5: The architecture of the Decoupled Detection Head, featuring parallel streams for classification and regression.

Figure 6: Visual samples from the constructed Copper Tube Defect Dataset.

\toprule Model	P	AP 50	AP 95	AP 75	mIoU	F 1	F 0.5	Avg
\midrule Cascade R-CNN [ 30 ]	0.594	0.732	0.479	0.334	0.755	0.661	0.619	0.596
Faster R-CNN [ 29 ]	0.500	0.746	0.471	0.265	0.613	0.537	0.737	0.553
CenterNet [ 49 ]	0.853	0.667	0.404	0.221	0.753	0.720	0.798	0.631
ATSS [ 50 ]	0.832	0.483	0.402	0.081	0.706	0.638	0.471	0.516
RetinaNet [ 32 ]	0.448	0.684	0.445	0.152	0.704	0.568	0.489	0.499
Dynamic R-CNN [ 51 ]	0.547	0.695	0.450	0.259	0.731	0.644	0.582	0.558

Table 1: Performance comparison with evaluated detection methods on CTDD. The best result is highlighted in bold, and the second-best result is underlined.

Figure 7: Visualization of detection results compared against the Ground Truth (GT). The last column shows the GT annotations, while the second to last column shows the predictions of our proposed method. Preceding columns display results from other models for comparison

\toprule Model	AP 50	AP 50-95	F 1	P	R
\midrule YOLOv11n	0.786	0.468	0.712	0.774	0.660
CF-YOLO	0.794	0.465	0.732	0.760	0.707
\bottomrule

Table 2: External validation results on NEU-DET.

Figure 8: Visual ablation analysis of key components. The left panel demonstrates the effectiveness of CPAM in localizing camouflaged defects, while the right panel illustrates the capability of FARM in capturing micro-scale anomalies. ’GT’ denotes Ground Truth.

\toprule CPAM	FARM	P	F 1	AP 50
\midrule -	-	0.843	0.756	0.749
$\checkmark$	-	0.842	0.783	0.819
-	$\checkmark$	0.838	0.780	0.800
$\checkmark$	$\checkmark$	0.882	0.796	0.823
\bottomrule

Table 3: Ablation study on the effectiveness of proposed components.

\toprule Variant	AP 50	AP ${}_{50\text{-}95}$	F 1	Precision
\midrule Full FARM (Ours)	0.823	0.488	0.796	0.882
No spatial branch	0.772	0.467	0.748	0.825
No channel branch	0.785	0.476	0.765	0.825
Conv-only	0.775	0.472	0.752	0.815
\bottomrule

Table 4: FARM mixer variant ablation on CTDD. Full FARM is the unmodified configuration used in CF-YOLO.

\toprule Configuration	AP 50	AP ${}_{50\text{-}95}$	F 1
\midrule FARM in neck (Ours)	0.823	0.488	0.796
FARM before neck	0.760	0.469	0.738
FARM after neck	0.770	0.474	0.748
\bottomrule

Table 5: FARM placement analysis on CTDD.

\toprule Configuration	AP 50	AP ${}_{50\text{-}95}$	F 1
\midrule CPAM mid-backbone (Ours)	0.823	0.488	0.796
CPAM shallow	0.730	0.462	0.705
CPAM deep	0.718	0.460	0.695
Fewer CPAM blocks (2 $\times$ )	0.770	0.468	0.745
\bottomrule

Table 6: CPAM depth and quantity analysis on CTDD.

\toprule $K_{S}$	$K_{L}$	P	F 1	AP 50
\midrule 3	3	0.823	0.703	0.751
3	5	0.854	0.786	0.804
3	7	0.846	0.785	0.814
5	7	0.564	0.494	0.462
5	11	0.850	0.784	0.809
11	11	0.844	0.744	0.783

Table 7: Ablation study on CPAM kernel settings.

K_{L}

denotes the Large-Kernel Perception branch, and

K_{S}

denotes the Small-Kernel Aggregation branch.

Explore similar work

Continuity-Driven Representation Learning for Industrial Defect Detection
Aug 18, 2026Minjong Kim, Hyun Jun Kim, Jeongrae Kim +2Object Detection Industrial Anomaly Detection
Chung-Ang University Seoul, Republic of Korea
Contrastive Augmented Transformer with Domain-specific Enhancement for Robust Multi-scenario Metal Surface Defect Detection
Jun 1, 2026Yiyao Liu, Wenxiao He, Liyuan Ren +1Contrastive Learning Data Augmentation
Glasgow College, University of Electronic Science and Technology of China, Chengdu, 611731, China · Institute of Applied Physics and Computational Mathematics, Beijing, China · Department of Industrial Engineering, Tsinghua University, Beijing, China
Template-Based Feature Aggregation Network for Industrial Anomaly Detection
Mar 24, 2026Wei Luo, Haiming Yao, Wenyong YuIndustrial Anomaly Detection Reconstruction-Based Anomaly Detection
State Key Laboratory of Precision Measurement Technology and Instruments, Tsinghua University, Beijing, 100084, China · State Key Laboratory of Digital Manufacturing Equipment and Technology, Huazhong University of Science and Technology, Wuhan, 430074, China

Aug 18, 2026J/K move · Enter open · S save

Continuity-Driven Representation Learning for Industrial Defect Detection

Minjong Kim, Hyun Jun Kim, Jeongrae Kim, Heeseung Shin+1

Chung-Ang University Seoul, Republic of Korea

Object Detection Industrial Anomaly Detection Visual Representation Learning

Open paper AskPDF arXiv

Industrial defect detection differs from natural-image object detection because inspection images are captured under controlled conditions and contain large normal-dominant regions with repetitive structures. Defects therefore appear as localized disruptions of otherwise predictable patterns, while conventional detectors rely mainly on sparse bounding-box supervision, resulting in weakly constrained normal-region representations. We propose a continuity-driven representation regularization framework that exploits normal-dominant regions as dense auxiliary supervision. The framework introduces two detector-agnostic objectives: Multi-Continuity Loss, which combines 1D patch-sequence prediction and 2D masked spatial prediction, and Differencing Loss, which regularizes first-order feature variation and second-order curvature between neighboring patch embeddings. Both objectives are applied with box-derived region weighting to stabilize normal-region representations while preserving defect-related discontinuities. Experiments on two real-world industrial datasets and the public NEU-DET benchmark, using six detector architectures including YOLO-family models, MambaYOLO, and DETR, demonstrate consistent improvements over native detector baselines. In the full-data setting, the proposed regularizers improve average [email protected]:0.95 by up to 3.49 percentage points on Industrial Metal, 5.38 percentage points on MEA, and 5.03 percentage points on NEU-DET. Under limited-data conditions, the gains become more pronounced, with Differencing Loss achieving improvements of up to 21.07 percentage points in [email protected] and 8.23 percentage points in [email protected]:0.95 on NEU-DET using only 25% of the training data. These results suggest that continuity-driven regularization provides an effective prior for improving industrial defect detection, particularly when annotated data are scarce.

Abstract

Explore similar work

Continuity-Driven Representation Learning for Industrial Defect Detection

Contrastive Augmented Transformer with Domain-specific Enhancement for Robust Multi-scenario Metal Surface Defect Detection

Template-Based Feature Aggregation Network for Industrial Anomaly Detection