A stronger score on the main test
An artificial-intelligence model designed to find tiny defects that blend into copper tubing scored higher than YOLOv11n on the study's main test, according to an arXiv preprint. On the Copper Tube Defect Dataset, known as CTDD, CF-YOLO recorded an AP50 of 0.823, up from 0.801 for YOLOv11n. Its F1-score was 0.796 versus 0.771, and its Precision was 0.882 versus 0.843. The figures are direct benchmark comparisons, so they show how the two systems performed in this test setting.
The model keeps a YOLOv11-based detector at its core and adds two modules. CPAM, short for Context Perception and Aggregation Module, is used for macro-texture context and local boundary aggregation. FARM, or Feature Aggregation and Refinement Module, is used to refine fine-grained anomalies with a linear-complexity operation. Together, they are presented as a real-time detector aimed at camouflaged industrial micro-defects.
CTDD is a manually annotated benchmark with 1,847 images and 4,898 bounding-box defect instances from copper-tube inspection scenarios. The images were randomly split into training, validation and testing sets in an 8:1:1 ratio. Models were trained from scratch for 200 epochs with batches of 16 images. Mosaic and Mixup augmentation were applied during training and switched off for the final 10 epochs.
AP50 is average precision measured at an intersection-over-union threshold of 0.5, using that box-overlap cutoff. The evaluation also included AP75, AP95, F0.5-score and box-level mIoU. Precision was defined from true-positive and false-positive predictions, while F1 combines Precision and Recall. The study therefore reports several views of detector performance rather than one score.
The gain was broad, not universal
On the CTDD comparison, CF-YOLO led six of the listed measures: Precision, AP50, AP75, F1-score, F0.5-score and Avg. It did not top every metric. ETDNet had the highest AP95, while RF-DETR had the highest mIoU. The reported pattern is a broad lead across the benchmark's measures, not a clean sweep.
Component tests pointed in the same direction. The full CPAM-plus-FARM configuration had the highest reported component-ablation performance and exceeded the ablation baseline by 3.9% in Precision, 4.0% in F1-score and 7.4% in AP@50. Separate design tests favored putting FARM inside the neck and CPAM in the mid-backbone; reducing the number of CPAM blocks lowered reported accuracy. The review notes that baseline metrics differ between the main comparison and the ablation tables, so the percentage gains belong to the ablation setup.
External data tells a more mixed story
Results on the external NEU-DET data were more mixed. The six original defect categories were merged into one class, producing a reported setting with 1,800 images and 4,189 defect instances. Its split contained 1,260 training images, 270 validation images and 270 test images, using seed 0. Against YOLOv11n, CF-YOLO scored 0.794 versus 0.786 on AP50 and 0.732 versus 0.712 on F1. It also had higher Recall, 0.707 versus 0.660, but lower Precision, 0.760 versus 0.774, and slightly lower AP50-95, 0.465 versus 0.468.
That split result tempers the CTDD comparison. The authors describe NEU-DET as preliminary external validation, not evidence that the model will generalize universally. The external test also cannot answer how performance varies across the six original categories, because they were merged before evaluation. CTDD itself is a single-class bounding-box benchmark assessed with a random split from one dataset. The evidence therefore supports comparative performance claims in the reported settings, while leaving broader cross-dataset behavior unresolved.
What the benchmark can and cannot say
The framework is trained end-to-end with a multi-task loss that jointly optimizes deterministic and probabilistic outputs. That design sits alongside the two refinement modules, but the benchmark does not by itself establish reductions in production errors, costs or quality failures. The supplied analysis also reports no measured latency, throughput or edge-device performance. Added context aggregation and feature refinement may make deployment harder on resource-constrained devices.
The record identifies the document as an arXiv preprint, version 1, dated 28 August 2026. No funding source is reported in the supplied text. The findings are best read as an early benchmark report: stronger on CTDD, uneven on NEU-DET, and still in need of testing across acquisition environments and deployment hardware.
Paper data and sources
Original title: CF-YOLO: Context-Aware Feature Refinement for Camouflaged Industrial Micro-Defect Detection
Authors: Xinda Yu, Kunxin Zheng, Chunan Yu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text