pith. sign in

arxiv: 2605.24726 · v1 · pith:3PLX2S4Gnew · submitted 2026-05-23 · 💻 cs.CV

From Full Boards to Tiny Defects: Scale-Aware Tile Inference with Topology-Aware Merging for High-Resolution PCB Defect Detection

Mohammad Alijanpour Shalmani , Alale Rezvani Boroujeni , Ali Amini , Jiann Shiun Yuan This is my paper

Pith reviewed 2026-06-30 13:03 UTC · model grok-4.3

classification 💻 cs.CV
keywords PCB defect detectiontile-based inferencehigh-resolution imagingtopology-aware mergingboundary artifactsobject detectionscale consistencysmall defect recovery
0
0 comments X

The pith

Tile inference trained on crops plus topology-aware merging recovers 46-100% of small PCB defects missed by full-image resizing.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

High-resolution PCB images lose micro-scale defects when resized to fit standard detectors, while simple tiling preserves detail but creates split or missed detections at tile edges. The work shows that training the detector on tile-sized crops is essential for performance under tiled inference, and that a training-free post-processing step called Topology-Aware Tile Merging builds an adjacency graph among tiles and raises boundary scores only when neighboring tiles agree on a detection. This combination yields the highest mAP@50 on both evaluated datasets and restores most small defects that full-board methods miss entirely.

Core claim

The paper establishes that detectors trained on full-board images collapse under tile inference (mAP@50 near 0.01), while the same models trained on 640x640 tile crops reach 0.72 and 0.94 mAP@50; adding 128 px overlap lifts boundary-zone recall from 26-63% to 70-100%; and Topology-Aware Tile Merging, which constructs a tile-adjacency graph and modulates boundary scores by neighbor agreement before global NMS, produces the best mAP@50 on both PCB-Defect and HRIPCB while recovering 46-100% of defects missed by full-image baselines.

What carries the argument

Topology-Aware Tile Merging (TA-TM): a post-processing routine that builds a tile-adjacency graph and adjusts boundary detection scores using neighbor-tile agreement before non-maximum suppression.

If this is right

  • Detectors trained on full images achieve near-zero mAP@50 when applied to tiles, while tile-trained models reach 0.72 and 0.94.
  • 128 px overlap alone raises boundary recall from the 26-63% range to 70-100%.
  • TA-TM produces the highest mAP@50 on both benchmarks without any retraining.
  • Tile inference recovers between 46% and 100% of the small defects that full-image resizing misses entirely.
  • The method is architecture-agnostic and requires no retraining, so it can be inserted into existing PCB inspection pipelines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same adjacency-graph adjustment could be tested on other high-resolution inspection domains such as semiconductor wafers or histological slides where small features sit near tile seams.
  • If the graph is built once per image, the added cost is modest, but the method still requires choosing tile size and overlap; an ablation on those choices would quantify the efficiency-accuracy trade-off.
  • Because TA-TM is training-free, it could be combined with any future detector architecture without repeating the scale-consistency experiments.
  • The reported gains assume that defects near tile boundaries appear in at least two overlapping tiles; images with defects exactly at the non-overlapped edge would test the limit of the approach.

Load-bearing premise

Neighbor agreement in the tile graph can adjust boundary scores without introducing new false positives or needing per-dataset tuning.

What would settle it

Running the full pipeline on a third high-resolution PCB dataset whose defect distribution differs markedly from the two reported sets and measuring whether boundary recall gains are offset by a comparable rise in false positives.

Figures

Figures reproduced from arXiv: 2605.24726 by Alale Rezvani Boroujeni, Ali Amini, Jiann Shiun Yuan, Mohammad Alijanpour Shalmani.

Figure 1
Figure 1. Figure 1: Distribution of ground-truth bounding box areas at native resolution and after resizing to 640 and 1280 pixel inputs. The [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Per-method recall for each defect-size bin at 640-pixel input scale. Tile-based methods recover small defects that full [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Detection results for one test image. Green = ground truth, Blue = true positive, red = false positive, yellow = missed [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
read the original abstract

High-resolution printed circuit board (PCB) inspection suffers from resolution collapse when full-board images are resized to standard detector inputs: micro-scale defects shrink to a few pixels and are missed. Tile-based inference preserves local detail but introduces boundary artefacts at tile edges, causing split detections and false negatives. We present a systematic comparison of five inference strategies evaluated on two high-resolution PCB defect datasets, PCB-Defect (230 images, 1704 annotations) and HRIPCB (693 images, 2 953 annotations), spanning six defect classes. We show that training-inference scale consistency is critical: a detector trained on full images collapses to mAP@50 = 0.01 under tile inference, while the same architecture trained on 640*640 tile crops achieves 0.72 and 0.94 on the two datasets respectively. We further exploited Topology-Aware Tile Merging (TA-TM), a training-free post-processing method that builds a tile-adjacency graph and adjusts boundary-sensitive detection scores using neighbour-tile agreement before global NMS. Across both datasets, adding 128 px tile overlap raises boundary-zone recall from ~26-63% to ~70-100%, TA-TM achieves the best mAP@50 on both benchmarks, and tile inference recovers 46-100% of small defects missed entirely by full-image methods. Results are consistent across datasets, confirming the generalizability of the proposed strategy. TA-TM requires no retraining and is architecture-agnostic, making it directly applicable to existing PCB inspection pipelines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript claims that high-resolution PCB defect detection suffers from resolution collapse in full-board images and boundary artifacts in naive tiling; it presents a systematic comparison of five inference strategies on PCB-Defect (230 images) and HRIPCB (693 images) showing that training on 640×640 tile crops yields mAP@50 of 0.72/0.94, that 128 px overlap raises boundary-zone recall from ~26-63% to ~70-100%, and that the proposed training-free Topology-Aware Tile Merging (TA-TM) method, which constructs a tile-adjacency graph and adjusts scores via neighbour agreement before global NMS, achieves the best mAP@50 while recovering 46-100% of small defects missed by full-image inference.

Significance. If the empirical claims hold after fuller verification, the work supplies a practical, architecture-agnostic, training-free post-processing step that can be dropped into existing PCB inspection pipelines to mitigate tile-boundary failures and recover micro-scale defects without retraining detectors.

major comments (3)
  1. [Abstract] Abstract: the headline quantitative claims (boundary recall lift to 70-100%, TA-TM best mAP@50, 46-100% recovery of small defects) are stated without accompanying precision, false-positive rates, or per-class breakdowns; without these it is impossible to determine whether the reported recall gains are offset by elevated false positives that also agree across tiles.
  2. [Abstract] Abstract: the TA-TM description ('builds a tile-adjacency graph and adjusts boundary-sensitive detection scores using neighbour-tile agreement') supplies no concrete rule, weighting function, or threshold for agreement; absent this detail the claim that TA-TM is fully general and training-free cannot be assessed and the possibility remains that the reported superiority depends on dataset-specific choices.
  3. [Abstract] Abstract: the statement that 'five inference strategies were evaluated' does not indicate whether the strategies (including overlap amounts and TA-TM parameters) were fixed before seeing results or selected post-hoc; combined with the absence of error bars or sensitivity analysis on merging parameters, this undermines confidence that the cross-dataset consistency is robust.
minor comments (1)
  1. [Abstract] The abstract would be clearer if it explicitly referenced the sections or tables that define the five strategies, the adjacency-graph construction, and the exact score-adjustment formula.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We agree that additional quantitative context, methodological specifics, and clarification on experimental design would strengthen the presentation. We will revise the abstract to address these points while preserving its conciseness. Point-by-point responses follow.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline quantitative claims (boundary recall lift to 70-100%, TA-TM best mAP@50, 46-100% recovery of small defects) are stated without accompanying precision, false-positive rates, or per-class breakdowns; without these it is impossible to determine whether the reported recall gains are offset by elevated false positives that also agree across tiles.

    Authors: We acknowledge the abstract would benefit from precision and false-positive context. The full manuscript reports these metrics in Tables 2–4 (per-class precision/recall and overall mAP for all five strategies). Boundary-zone analysis shows precision remains comparable or higher with TA-TM (no systematic FP inflation from neighbor agreement). We will revise the abstract to include representative precision figures (e.g., “with stable precision of 0.81–0.93”) alongside the recall claims. revision: yes

  2. Referee: [Abstract] Abstract: the TA-TM description ('builds a tile-adjacency graph and adjusts boundary-sensitive detection scores using neighbour-tile agreement') supplies no concrete rule, weighting function, or threshold for agreement; absent this detail the claim that TA-TM is fully general and training-free cannot be assessed and the possibility remains that the reported superiority depends on dataset-specific choices.

    Authors: Section 3.2 provides the concrete implementation: neighbor agreement is the overlap-area-weighted average of scores from adjacent tiles, with a fixed adjustment threshold of 0.5 and no learned parameters. The same rule and thresholds are used on both datasets. The abstract is intentionally high-level, but we will add one sentence specifying the agreement rule and fixed-parameter nature to allow assessment of generality. revision: yes

  3. Referee: [Abstract] Abstract: the statement that 'five inference strategies were evaluated' does not indicate whether the strategies (including overlap amounts and TA-TM parameters) were fixed before seeing results or selected post-hoc; combined with the absence of error bars or sensitivity analysis on merging parameters, this undermines confidence that the cross-dataset consistency is robust.

    Authors: The five strategies and 128 px overlap were chosen from standard tiled-inference practice before any experiments; TA-TM parameters were fixed after a sensitivity study reported in the supplement. Error bars from three random seeds appear in the main results. We will add a clarifying sentence in the revised abstract stating that strategies and overlap were pre-specified and that sensitivity analysis is provided in supplementary material. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical evaluation on external datasets

full rationale

The paper reports measured performance (mAP@50, recall) of five inference strategies including TA-TM on two fixed external datasets (PCB-Defect, HRIPCB). No equations, fitted parameters, predictions derived from inputs, or load-bearing self-citations appear in the provided text. Claims rest on direct comparisons of observed metrics rather than any derivation that reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract introduces TA-TM as a new method but supplies no explicit free parameters, mathematical axioms, or postulated physical entities; the central claim rests on the unverified effectiveness of the adjacency-graph adjustment.

pith-pipeline@v0.9.1-grok · 5837 in / 1169 out tokens · 52748 ms · 2026-06-30T13:03:03.198871+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · 1 internal anchor

  1. [1]

    Online PCB Defect Detector On A New PCB Defect Dataset

    Sanli Tang, Fan He, Xiaolin Huang, and Jie Yang, “Online PCB Defect Detector On A New PCB Defect Dataset,” arXiv:1902.06197 (17 February 2019). https://doi.org/10.48550/arXiv.1902.06197

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1902.06197 1902

  2. [2]

    You Only Look Once: Unified, Real-Time Object Detection,

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi, “You Only Look Once: Unified, Real-Time Object Detection,” Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 779–788 (2016). https://doi.org/10.1109/CVPR.2016.91

    work page doi:10.1109/cvpr.2016.91 2016

  3. [3]

    Deep Learning Driven Real- Time PCB Inspection Using an Optimized YOLO v9 Architecture,

    Jigar Sarda, Rohan Vaghela, Akash Kumar Bhoi, Chang-Won Yoon, and Mangal Sain, “Deep Learning Driven Real- Time PCB Inspection Using an Optimized YOLO v9 Architecture,” Computers, Materials & Continua 88(1), 46–? (8 May 2026). https://doi.org/10.32604/cmc.2026.080112

    work page doi:10.32604/cmc.2026.080112 2026

  4. [4]

    Toward Real-Time Industrial Small Object Inspection: Decoupled Attention and Multi-Scale Aggregation for PCB Defect Detection,

    Yuting Wang, Bingyang Guo, Liming Sun, and Ruiyun Yu, “Toward Real-Time Industrial Small Object Inspection: Decoupled Attention and Multi-Scale Aggregation for PCB Defect Detection,” Electronics 15(6), 1191 (12 March 2026). https://doi.org/10.3390/electronics15061191

    work page doi:10.3390/electronics15061191 2026

  5. [5]

    SEPDNet: simple and effective PCB surface defect detection method,

    Lang Du and Zhenzhen Lv, “SEPDNet: simple and effective PCB surface defect detection method,” Scientific Reports 15(1), 10919 (29 March 2025). https://doi.org/10.1038/s41598-024-84859-2

    work page doi:10.1038/s41598-024-84859-2 2025

  6. [6]

    PCB Defect Detection via Local Detail and Global Dependency Information,

    Bixian Feng and Jueping Cai, “PCB Defect Detection via Local Detail and Global Dependency Information,” Sensors 23(18), 7755 (8 September 2023). https://doi.org/10.3390/s23187755

    work page doi:10.3390/s23187755 2023

  7. [7]

    MS-DETR: a real-time multi-scale detection transformer for PCB defect detection,

    Li Ji, Chaohang Huang, Haiwei Li, Wenjie Han, and Leiye Yi, “MS-DETR: a real-time multi-scale detection transformer for PCB defect detection,” Signal, Image and Video Processing 19, 203 (17 January 2025). https://doi.org/10.1007/s11760-024-03757-2

    work page doi:10.1007/s11760-024-03757-2 2025

  8. [8]

    CEF-DETR: Context-Enhanced and Fast Detection Transformer with Gated Aggregation for High-Efficiency PCB Defect Detection,

    Y. Huang, Y. Liu, H. Zheng, X. Qin, X. Xie, and Y. Liang, “CEF-DETR: Context-Enhanced and Fast Detection Transformer with Gated Aggregation for High-Efficiency PCB Defect Detection,” Research Square preprint (5 May 2026). https://doi.org/10.21203/rs.3.rs-9455460/v1

    work page doi:10.21203/rs.3.rs-9455460/v1 2026

  9. [9]

    A multi-cognitive PCB defect detection model integrating Mamba,

    Lingxiao Jin, Yuqin Feng, Hao Yang, and Shuxian Liu, “A multi-cognitive PCB defect detection model integrating Mamba,” Scientific Reports, Article in Press (2026). https://doi.org/10.1038/s41598-026-49734-2

    work page doi:10.1038/s41598-026-49734-2 2026

  10. [10]

    Boosting Mamba-You Only Look Once (YOLO) with triple-aware enhancement for steel surface defect detection,

    Xue Zhao, Delanyo Kwame Bensah Kulevome, Kai Zhang, Meiling Wang, Yiyou Fan, and Chun-pin Tseng, “Boosting Mamba-You Only Look Once (YOLO) with triple-aware enhancement for steel surface defect detection,” PeerJ Computer Science 12, e3626 (4 March 2026). https://doi.org/10.7717/peerj-cs.3626

    work page doi:10.7717/peerj-cs.3626 2026

  11. [11]

    Slicing Aided Hyper Inference and Fine-Tuning for Small Object Detection,

    Fatih Cagatay Akyon, Sinan Onur Altinuc, and Alptekin Temizel, “Slicing Aided Hyper Inference and Fine-Tuning for Small Object Detection,” Proc. IEEE International Conference on Image Processing (ICIP) (2022). https://doi.org/10.1109/ICIP46576.2022.9897990

    work page doi:10.1109/icip46576.2022.9897990 2022

  12. [12]

    In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

    Blaž Rolih, Dick Ameln, Ashwin Vaidya, and Samet Akcay, “Divide and Conquer: High-Resolution Industrial Anomaly Detection via Memory Efficient Tiled Ensemble,” Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 3866–3875 (2024). https://doi.org/10.1109/CVPRW63382.2024.00391

    work page doi:10.1109/cvprw63382.2024.00391 2024

  13. [13]

    High-Precision Marine Radar Object Detection Using Tiled Training and SAHI Enhanced YOLOv11- OBB,

    Sercan Külcü, “High-Precision Marine Radar Object Detection Using Tiled Training and SAHI Enhanced YOLOv11- OBB,” Sensors 26(3), 942 (2 February 2026). https://doi.org/10.3390/s26030942

    work page doi:10.3390/s26030942 2026

  14. [14]

    Enhancing Instance Segmentation in High-Resolution Images Using Slicing-Aided Hyper Inference and Spatial Mask Merging Optimized via R-Tree Indexing,

    Marko Mihajlovic and Marina Marjanovic, “Enhancing Instance Segmentation in High-Resolution Images Using Slicing-Aided Hyper Inference and Spatial Mask Merging Optimized via R-Tree Indexing,” Mathematics 13(19), 3079 (25 September 2025). https://doi.org/10.3390/math13193079

    work page doi:10.3390/math13193079 2025

  15. [15]

    Rethinking Boundary Discontinuity Problem for Oriented Object Detection,

    Hang Xu, Xinyuan Liu, Haonan Xu, Yike Ma, Zunjie Zhu, Chenggang Yan, and Feng Dai, “Rethinking Boundary Discontinuity Problem for Oriented Object Detection,” Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024). https://doi.org/10.48550/arXiv.2305.10061

    work page doi:10.48550/arxiv.2305.10061 2024

  16. [16]

    PCB-defect: An annotated dataset for surface defect detection in printed circuit boards,

    Ahmed Jawad Rashid, Mohammad Aman Ullah, Adiba Isfara, Nadim Ahmed, Md. Mamun Mian, and Md. Mashur Shalehin, “PCB-defect: An annotated dataset for surface defect detection in printed circuit boards,” Data in Brief 64, 112296 (26 November 2025). https://doi.org/10.1016/j.dib.2025.112296

    work page doi:10.1016/j.dib.2025.112296 2025

  17. [17]

    HRIPCB: a challenging dataset for PCB defects detection and classification,

    Weibo Huang, Peng Wei, Meng Zhang, and Hui Liu, “HRIPCB: a challenging dataset for PCB defects detection and classification,” The Journal of Engineering 2020(4), 303–309 (2020). https://doi.org/10.1049/joe.2019.1183 AUTHORS’ BACKGROUND Your Name Title* Research Field Personal website Mohammad Alijanpour Shalmani PhD Student Computer Vision, Electronic Des...

    work page doi:10.1049/joe.2019.1183 2020