BoardVision: Motherboard Defect Detection

• The paper introduces BoardVision, a reproducible framework that fuses YOLOv7 and Faster R-CNN using Confidence-Temporal Voting (CTV) for detecting mounting- and wiring-related motherboard defects.
• It leverages the MiracleFactory Motherboard Dataset with 389 high-resolution images and 2,860 annotated defect instances across 11 classes to benchmark performance under realistic conditions.
• The CTV ensemble harmonizes high-speed detection and high-recall error correction, enhancing precision, recall, and stability against imaging perturbations in production QA.

BoardVision is a reproducible framework for robust assembly-level motherboard defect detection that consolidates two leading computer vision detectors—YOLOv7 and Faster R-CNN—through a lightweight, interpretable ensemble strategy termed Confidence-Temporal Voting (CTV). It addresses the underexplored challenge of detecting mounting- and wiring-related defects in assembled motherboards under realistic conditions, providing a practical and deployable end-to-end system that bridges the research-production gap in electronics quality assurance (Hill et al., 16 Oct 2025).

1. Dataset and Task Scope

BoardVision’s empirical foundation is the MiracleFactory Motherboard Dataset, which comprises 389 high-resolution (640×640) RGB images with a total of 2,860 defect instances annotated across 11 assembly-level defect classes. These classes capture a representative spectrum of faults encountered during motherboard production, including missing screws, loose or incorrect fan fittings, surface scratches, and detached connectors.

Classes and Distribution

Class Name	Instance Count
Screws	806
CPU_FAN_Screws	685
CPU_FAN_NO_Screws	326
CPU_fan	313
No_Screws	196
CPU_fan_port	159
CPU_FAN_Screw_loose	99
Scratch	95
Incorrect_Screws	63
CPU_fan_port_detached	60
Loose_Screws	58

Defect types of interest include missing screws, loose/incorrect mounting, fan mis-wiring, connector detachment, and surface scratches—reflecting real-world assembly line QA requirements.

2. Baseline Detector Performance

BoardVision benchmarks two primary object detectors:

• YOLOv7: a real-time, single-stage CNN detector, trained at 640×640 input resolution, optimized with SGD at a learning rate of 0.01 (batch size ≈ 16).
• Faster R-CNN (ResNet-50+FPN): a two-stage region proposal network with a 0.005 learning rate, identical batch size, and early stopping on validation mAP.

Training and evaluation leverage PyTorch (v2.0), CUDA (11.7), and a GTX 1080 GPU. The validation protocol uses a held-out 45-image test set. In addition, data augmentation at inference simulates robustness factors, including horizontal flips, sharpness modulation (Gaussian unsharp masking), and linear brightness variations.

Metric	YOLOv7	Faster R-CNN
[email protected]	0.914	0.766
[email protected]:0.95	0.606	0.495
Precision	0.964	0.953
Recall	0.956	0.718
F1-score	0.960	0.819
FPS	22–25	8–10

YOLOv7 achieves superior inference speed and precision, though Faster R-CNN identifies defects missed by YOLO, particularly rare or ambiguous cases.

3. Confidence-Temporal Voting (CTV) Ensemble

The core technical contribution is the CTV ensemble, designed to harmonize the high-precision, high-speed characteristics of YOLOv7 with the high-recall, error-correcting nature of Faster R-CNN.

Detection Fusion Methodology

For each frame $t$, let detections be denoted $(b_i^{(t)}, c_i^{(t)}, p_i^{(t)})$. YOLO and Faster R-CNN detections are greedily matched by class and IoU threshold ($t_{IoU}$). For each matched pair $(Y_i, R_j)$ of class $c$:

$$S_Y = (p_Y)^\gamma \, F1_{\text{YOLO},c},\quad S_R = (p_R)^\gamma \, F1_{\text{FRCNN},c}$$

Fused box coordinates:

$$b^* = \frac{S_Y\,b_Y + S_R\,b_R}{S_Y + S_R},\quad p^* = \max(p_Y,p_R),\quad c^* = c$$

Unmatched (solo) detections are accepted based on interpretable rules: (1) high confidence ($p \ge \mathit{solo\_strong}$), (2) superior per-class F1 and thresholded confidence, (3) near-tie in F1 and $p \ge 0.95$. Class-wise NMS is lastly applied.

A temporal voting function for sequence-level smoothing is:

$$D_i^{(T)} = \begin{cases} 1, & \sum_{t=1}^T \mathbf{1}(p_i^{(t)} > \tau) \ge V \ 0, & \text{otherwise} \end{cases}$$

Pseudocode Outline

1. Load YOLO and FRCNN detections for each frame.
2. Match boxes by class and IoU.
3. Compute weighted box fusion.
4. Apply solo rules for unmatched detections.
5. Perform NMS and output.

4. Robustness to Realistic Imaging Perturbations

The BoardVision evaluation protocol stresses resilience under deployment-relevant transformations:

• Flip (horizontal mirror),
• Sharpness Up (unsharp mask),
• Brightness Up/Down (linear adjustment).

Average and standard deviation over these perturbations provide insight into both mean performance and stability.

Metric	YOLOv7 (mean ± std)	Faster R-CNN	CTV Ensemble
Precision	0.962 ± 0.006	0.954 ± 0.008	0.964 ± 0.006
Recall	0.949 ± 0.009	0.713 ± 0.003	0.954 ± 0.009
F1-score	0.958 ± 0.007	0.816 ± 0.005	0.957 ± 0.006

YOLOv7 exhibits minor F1 drops under sharpness increase (to 0.942). Faster R-CNN demonstrates greater sensitivity across all perturbations (mean F1 = 0.816). The CTV ensemble both restores YOLOv7’s recall and reduces F1-score variance by approximately 15%, indicating robust stability under adverse imaging conditions. CTV’s mean robustness score $(b_i^{(t)}, c_i^{(t)}, p_i^{(t)})$0 is marginally lower than YOLOv7’s, but its variance is substantially reduced.

5. Operator-centric Deployment: GUI-driven Inspection Tool

BoardVision incorporates an operator-facing PySide6/Qt GUI for real-time inspection. The interface presents three synchronized panes (YOLOv7, Faster R-CNN, and CTV outputs), provides selectable input sources (files, webcam, RTSP streams), and exposes parameters ($(b_i^{(t)}, c_i^{(t)}, p_i^{(t)})$1, $(b_i^{(t)}, c_i^{(t)}, p_i^{(t)})$2, solo thresholds) alongside runtime controls (frame skip, pause/resume). Bounding boxes are color-coded by originating model and decision status; operator corrections are logged for later audit.

The deployment stack features:

• Preprocessing and model loading (CPU/GPU)
• Parallel YOLOv7 & Faster R-CNN inference per frame
• Immediate CTV fusion pipeline execution
• Live update and overlay of fused boxes
• Audit log of operator input
• Modes for both streaming (real-time QA) and file-based (batch) operation

This integration couples model transparency, human oversight, and system usability in a production QA context.

6. Quantitative and Practical Findings

The CTV ensemble achieves the highest overall performance equilibrium: [email protected] = 0.921, [email protected]:0.95 = 0.604, Precision = 0.967, Recall = 0.962, F1 = 0.964, at inference speed comparable to Faster R-CNN alone (8–10 FPS). Notably, rare classes such as No_Screws realize marked F1 improvement (0.926 [YOLOv7] to 0.963 [CTV]). Robustness analysis demonstrates that CTV stabilizes error rates under lighting and sharpness variations, reducing F1 variance by approximately 15%.

Parameter sweeps reveal the system is stable for $(b_i^{(t)}, c_i^{(t)}, p_i^{(t)})$3, $(b_i^{(t)}, c_i^{(t)}, p_i^{(t)})$4, and $(b_i^{(t)}, c_i^{(t)}, p_i^{(t)})$5; more extreme parameterizations trade off between precision and recall.

Findings underscore that ensemble frameworks leveraging explicit, interpretable rules (e.g., high-confidence overrides, per-class F1-weighting) outperform complex black-box meta-learners in this domain, and GUI-driven transparency supports operator trust and auditability.

In sum, BoardVision illustrates a pathway from laboratory detectors and benchmarks to an operational, operator-interpretable QA system for assembly-level motherboard manufacturing, harmonizing speed, precision, recall, robustness, and human-facing deployment (Hill et al., 16 Oct 2025).