💻 computer science

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

This paper addresses the resolution collapse and boundary artifacts in high-resolution PCB defect detection by demonstrating that training detectors on tile crops and applying a novel, training-free Topology-Aware Tile Merging (TA-TM) post-processing method significantly improves small defect recall and overall detection accuracy compared to traditional full-image or standard tile inference strategies.

Original authors: Mohammad Alijanpour Shalmani, Alale Rezvani Boroujeni, Ali Amini, Jiann Shiun Yuan

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammad Alijanpour Shalmani, Alale Rezvani Boroujeni, Ali Amini, Jiann Shiun Yuan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find tiny scratches, missing screws, or broken wires on a massive, high-definition blueprint of a circuit board. This blueprint is so huge that if you shrink it down to fit on a standard computer screen (like resizing a giant mural to fit on a postage stamp), those tiny defects disappear completely. They become invisible specks.

This is the problem the paper tackles: How do you find microscopic defects on a giant image without losing the details?

Here is the story of their solution, broken down into simple concepts and analogies.

1. The Problem: The "Postage Stamp" Effect

Imagine you have a high-resolution photo of a forest. If you shrink that photo to the size of a postage stamp to fit it into a standard photo frame, you can no longer see individual leaves or small birds; they just blur into green or brown pixels.

In the world of circuit boards (PCBs), this is called "Resolution Collapse."

The Old Way: People took the whole giant board image, shrunk it down to fit the AI's "brain" (which expects a standard size), and tried to find defects.
The Result: The AI, trained to look at standard-sized images, couldn't see the tiny defects because they had shrunk to nothing. The AI essentially went blind to the very things it was supposed to find.

2. The First Fix: The "Puzzle Piece" Strategy

Instead of shrinking the whole picture, the researchers decided to cut the giant image into smaller, manageable pieces (tiles), like slicing a large pizza into slices.

The Strategy: They fed these smaller slices to the AI one by one. Since the slices weren't shrunk, the tiny defects remained big and clear.
The Catch: When you cut a picture into pieces, you create edges. If a defect sits right on the line where two slices meet, it gets chopped in half. The AI sees a half-defect in one slice and a half-defect in the next, gets confused, and might throw both away because they look "incomplete." This is called a boundary artifact.

3. The Second Fix: The "Overlapping" Strategy

To fix the chopped-up defects, the researchers made the puzzle pieces overlap.

The Analogy: Imagine laying down floor tiles, but instead of just touching edges, you slide them so they overlap by a few inches. Now, if a scratch is on the edge, it appears fully in the left tile and fully in the right tile.
The Result: This helped the AI see the defects clearly, but it created a new problem: The AI might see the same defect twice (once in the left tile, once in the right) and get confused about which one is the "real" one.

4. The Secret Sauce: "Topology-Aware Tile Merging" (TA-TM)

This is the paper's main innovation. It's a smart rulebook the computer uses after it has looked at all the tiles, but before it gives you the final answer.

How it works: Imagine the AI is a team of detectives.
- Detective A looks at the left tile and sees a "half-suspect" near the edge. They aren't 100% sure, so they give it a low confidence score.
- Detective B looks at the overlapping right tile and sees the other half of the same suspect. They also give it a low score.
- The Old Way (Standard AI): The computer would look at these two low-confidence reports and say, "Neither is strong enough, let's ignore them." The defect is missed.
- The New Way (TA-TM): The computer acts like a smart manager. It sees that Detective A and Detective B are looking at the same spot from different angles. It says, "Hey, if both of you see something there, even if you aren't sure individually, let's combine your notes and boost the confidence!"
- The Result: The computer realizes, "Ah, this is a real defect!" and keeps it.

5. The Big Discovery: Training Matters More Than the Tool

The paper found something surprising that changes how people should build these AI systems.

The Mismatch: If you train an AI by showing it whole, shrunk-down images, it learns to look for "tiny, blurry specks." If you then try to use that same AI on the "Puzzle Piece" strategy (where defects are big and clear), the AI fails miserably. It's like training a dog to hunt rabbits by showing it pictures of rabbits, then expecting it to hunt elephants.
The Fix: You must train the AI exactly the way you will use it. If you plan to use the "Puzzle Piece" strategy, you must train the AI on puzzle pieces.
The Proof: When they trained the AI on puzzle pieces, it became a master at finding defects. When they tried to use a "whole-image" trained AI on puzzle pieces, it failed completely (dropping from 72% success to 1% success).

Summary of Results

The "Whole Image" approach: Missed almost all tiny defects.
The "Puzzle Piece" approach: Found most defects but struggled with those on the edges.
The "Overlapping Puzzle" + "Smart Manager" (TA-TM): Found the most defects, including the tricky ones on the edges, without needing to retrain the AI or change its brain. It just needed a smarter way to combine the clues.

In short: To find tiny things on a giant image, don't shrink the image. Cut it into overlapping pieces, train your AI on those pieces, and use a smart "manager" to combine the clues from the edges. This simple change turned a failing system into a highly accurate one.

Technical Summary: Scale-Aware Tile Inference with Topology-Aware Merging for High-Resolution PCB Defect Detection

Problem Statement
High-resolution Printed Circuit Board (PCB) inspection faces a critical challenge known as "resolution collapse." Standard object detectors (e.g., YOLO-family) operate on fixed input sizes (typically 640×640 pixels). When full-board high-resolution images (often 1,500–6,000 pixels) are resized to fit these inputs, micro-scale defects shrink to a few pixels, rendering them undetectable. Conversely, while tile-based inference (slicing images into smaller crops) preserves local detail, naive application introduces boundary artifacts. Defects straddling tile edges are often partially visible, leading to split detections, false negatives, or low-confidence scores that are suppressed by standard Non-Maximum Suppression (NMS). Furthermore, a significant mismatch exists between training and inference scales: models trained on resized full images fail when applied to native-resolution tiles, and vice versa.

Methodology
The authors propose a framework that prioritizes training-inference scale consistency and introduces a novel post-processing merging strategy, rather than modifying detector architectures. The pipeline consists of four stages:

Scale-Aware Training: Instead of resizing full images, the detector is trained exclusively on 640×640 pixel crops extracted directly from high-resolution source images. A sliding window with a 512-pixel stride (128-pixel overlap) ensures defects near boundaries are fully visible during training. Ground-truth boxes are clipped, retaining only those with at least 40% visibility.
Inference Strategies: The study evaluates five strategies using the same tile-trained detector:
- Full-Image Inference: Resizing the entire board to 640 or 1280 pixels.
- Tile-Based (No Overlap): Processing non-overlapping tiles.
- Tile-Based (With Overlap): Processing tiles with 128-pixel overlap.
Topology-Aware Tile Merging (TA-TM): A training-free, architecture-agnostic post-processing method designed to resolve boundary artifacts.
- Graph Construction: The tile grid is modeled as an adjacency graph.
- Boundary Detection: For each prediction, the distance to the nearest tile edge is calculated. Detections within a threshold ( $\tau = 16$ px) are flagged as "boundary-sensitive."
- Agreement Scoring: For a boundary-sensitive detection, the algorithm checks adjacent tiles for complementary detections of the same class that overlap across the shared boundary.
- Score Adjustment: If a complementary detection exists, the confidence score of the boundary-sensitive detection is boosted ( $s' = \min(1.0, s + \lambda \cdot A)$ , where $\lambda=0.2$ ).
- Global NMS: All predictions are merged using class-aware NMS based on these adjusted scores, allowing split detections confirmed by neighbors to survive suppression.

Key Contributions

Resolution Collapse Analysis: The paper empirically demonstrates that training-inference scale consistency is the dominant factor in performance. A model trained on full images collapses to mAP@50 = 0.01 when applied to tile crops, whereas a model trained on tile crops achieves mAP@50 of 0.72–0.94.
Systematic Comparison: A rigorous evaluation of five inference strategies across two high-resolution datasets (PCB-Defect and HRIPCB) covering six defect classes.
TA-TM Algorithm: The introduction of a lightweight, post-processing method that improves boundary-zone recall without retraining or architectural changes.
Deployment-Oriented Metrics: Evaluation includes standard mAP alongside specific metrics for small-defect recall and boundary-zone recall, addressing practical industrial needs.

Results
Evaluated on PCB-Defect (230 images) and HRIPCB (693 images):

Scale Consistency: Tile-based inference with a tile-trained model significantly outperforms full-image inference. Full-640 inference yielded mAP@50 of 0.072 (PCB-Defect) and 0.192 (HRIPCB), while the proposed Tile+Overlap+TA-TM strategy achieved 0.722 and 0.936, respectively.
Small Defect Recovery: Tile-based methods recovered 46–100% of small defects (16–64 px² at 640-scale) that were entirely missed by full-image methods. On HRIPCB, tile inference achieved 100% recall for these small defects.
Boundary Artifact Suppression: Adding 128-pixel overlap increased boundary-zone recall (0–16 px from edge) from ~26% to ~70% on PCB-Defect and from ~63% to 100% on HRIPCB. TA-TM maintained or slightly improved these results while achieving the highest overall mAP.
Efficiency: While tile-based methods are slower (247–459 ms/image) than full-image inference (54–98 ms), the authors note this remains practical for offline industrial inspection given the substantial gains in accuracy.

Significance and Claims
The paper claims that the primary bottleneck in high-resolution PCB inspection is not the detector architecture (e.g., YOLO vs. Transformer) but the alignment of training and inference scales. The authors argue that practitioners should prioritize scale-aware training protocols before optimizing model complexity.

The proposed TA-TM method is presented as a generalizable, "plug-and-play" solution that can be applied to existing inspection pipelines without retraining. The authors modestly acknowledge that TA-TM offers moderate improvements when the base detector is already strong, suggesting its impact would be more pronounced in lower-SNR conditions or with weaker detectors. The work reframes high-resolution defect detection as an inference strategy problem, providing a systematic approach to handling resolution collapse and boundary artifacts in industrial computer vision.