PCB-MC: Missing Component Analysis in Printed Circuit Boards

I. INTRODUCTION

• Existing PCB benchmarks lack annotations for absent components at their expected locations, leaving missing-component detection comparatively underexplored.Unlike conventional object detection, the task relies on cues such as exposed solder pads and silkscreen outlines.Source: §I · p. 1
• PCB-MC extends RF100 with footprint-level labels across 615 images, 31 categories, over 9,400 missing instances, and 197 distinct board types.Each board type corresponds to a unique PCB layout, with an average of 3.12±1.02 samples per type.Source: §I · p. 1
• Figure 1 color-codes missing-component boxes by type and zooms in on footprint cues including exposed solder pads and silkscreen outlines.Capacitors are blue, resistors green, ICs yellow, and LEDs pink.Source: Figure 1 · p. 1
• Board type-aware splits prevent layout memorization, while 1024×1024 inputs preserve fine-grained structural cues for comparisons across supervised and unsupervised methods.The evaluation also examines candidate-region generation and classification under weak visual evidence.Source: §I · p. 1
• PCB-MC annotations and the evaluation protocol are slated for public release to support reproducible research on structural absence detection.The passage qualifies availability as “upon acceptance.”Source: §I · p. 1

A. Computer Vision in Industrial Inspection

• Anomaly-detection studies focus on learning normal patterns and detecting deviations,, but such methods target visible defects and are less effective with small components or limited training data.The passage identifies complex production lines as a challenging setting.Source: A. Computer Vision in Industrial… · p. 2

B. PCB Inspection and Defect Detection

• Low- and medium-volume production, including prototyping, specialized boards, repair, and small batches, often relies on human assembly or rework, motivating automated inspection for missing components.The passage connects these systems to preventing latent failures, reducing manual reinspection, and improving operator consistency.Source: B. PCB Inspection and Defect Detection · p. 2
• PCB inspection progressed from alignment- and illumination-sensitive template matching and image differencing to CNN and transformer detectors for component localization and defect detection –, –.Two-stage detectors have also been applied, with higher computational overhead reported for Faster R-CNN and Mask R-CNN –.Source: B. PCB Inspection and Defect Detection · p. 2
• Most existing PCB work detects visible components or surface defects; multi-class missing-component detection remains more challenging than single-defect scenarios and needs dedicated datasets and evaluation protocols.Examples of targeted advances include small-object detection and contextual reasoning.Source: B. PCB Inspection and Defect Detection · p. 2

C. PCB Datasets

• FPIC, FICS-PCB, PCB-Metal, and PCB analysis target component recognition; DeepPCB, PCB Defect, and DsPCBSD focus on visible anomalies or template comparison, and these datasets do not explicitly model missing components at expected locations.Source: C. PCB Datasets · p. 2

III. PCB-MC: A DATASET FOR MISSING COMPONENT DETECTION

• PCB-MC derives from RF100 and adds manually reviewed missing-component annotations, with corrected boxes, standardized labels, merged visually equivalent subclasses, and ambiguous categories removed.The protocol annotates absent components at their expected footprint locations and cross-validates annotations across multiple annotators.Source: §III · p. 2
• PCB-MC contains 615 images across 31 classes, including 23 present-component classes, eight of which also identify absent components; passive components dominate the annotations.The passage describes this imbalance as reflecting realistic assembly, where small passive components are frequent and prone to omission.Source: §III · p. 2
• Because missing components are small relative to full-board images, the dataset motivates high-resolution evaluation to preserve fine-grained footprint cues.This passage provides the spatial rationale for the evaluation setting.Source: §III · p. 2
• 117,091 present and 9,435 missing component instances are annotated at expected footprints using solder pads, silkscreen outlines, and layout regularities.Missing-component locations support supervised learning for absence localization.Source: §III · p. 2

IV. BENCHMARK PROTOCOL ON PCB-MC

• The paper defines tasks on the newly introduced PCB-MC dataset.The passage does not specify the task definitions.Source: §IV · p. 2
• The paper describes implemented methods for missing-object detection.The passage does not identify individual methods.Source: §IV · p. 2
• The paper describes a designed evaluation protocol for its dataset and detection methods.The passage does not provide protocol details.Source: §IV · p. 2

• Task A jointly detects present and missing components across all 31 categories in 615 images.Source: A. Benchmark Tasks · p. 2
• Task M restricts detection to the 8 missing-component classes across 293 images.Source: A. Benchmark Tasks · p. 2
• Task C restricts detection to the 23 present-component classes across 615 images.Source: A. Benchmark Tasks · p. 3
• These tasks enable controlled comparison of standard component detection and structural absence detection under consistent annotation conditions.Source: A. Benchmark Tasks · p. 3

• Five supervised detector architectures span YOLOv8/11/26, RT-DETR, and D-FINE, with tiled SAHI inference also evaluated under identical settings.Source: B. Missing Component Detection Methods · p. 3
• The two-stage pipeline uses YOLOv11 proposals generated at confidence 0.05, averaging approximately 240 candidate boxes per image.Source: B. Missing Component Detection Methods · p. 3
• A ResNet-18 classifier labels proposals as present or missing, with expanded patches capturing pad geometry and silkscreen context.Focal loss addresses the 2:1 present-to-missing class ratio.Source: B. Missing Component Detection Methods · p. 3
• The anomaly-detection comparison covers PatchCore, PaDiM, DRAEM, and Reverse Distillation, trained on complete boards and tested on boards with missing components.Source: B. Missing Component Detection Methods · p. 3
• Anomaly maps become boxes through 0.995-quantile thresholding, connected-component extraction with noise removal, and non-maximum suppression at IoU 0.3.Source: B. Missing Component Detection Methods · p. 3

• FNR is emphasized for missing-component classes alongside mAP@0.5 and F1 score.Source: C. Evaluation · p. 3
• High-resolution images and small components motivate applying missing-object detection methods at the image-patch level.Source: C. Evaluation · p. 3
• Board-type-aware 5-fold validation keeps all patches from a board type in one fold, using 70% of board types for training and 30% for validation.Results are mean ± standard deviation across folds, without task-specific hyperparameter tuning.Source: C. Evaluation · p. 3

• Models use COCO-pretrained weights and are fine-tuned for 300 epochs with AdamW, an initial learning rate of 1×10−4, and cosine scheduling.Source: D. Implementation Details · p. 3
• Augmentation includes flipping, rotation, brightness/contrast adjustment, and mosaic; inference uses a 0.25 confidence threshold and class-aware non-maximum suppression.Source: D. Implementation Details · p. 3
• Experiments use NVIDIA GPU acceleration at 1024 × 1024 resolution to preserve fine-grained cues such as pads and silkscreen outlines.Source: D. Implementation Details · pp. 3–4

Benchmark Task Performance

• On task M, the proposed two-stage approach achieves the highest mAP (0.15) and lowest FNR (0.75), outperforming one-stage detectors whose FNR values exceed 0.87.Its F1 (0.13) remains comparable to one-stage detectors.Source: §V · p. 4
• On task A, YOLOv11 reaches mAP (0.30), while YOLOv11+SAHI has the best F1 (0.47) and lowest FNR (0.50); on task C, YOLOv11 reaches mAP 0.42 and F1 0.50.FNR remains approximately 0.50–0.70 across methods on these tasks.Source: §V · p. 4

Anomaly Detection vs. Object Detection

• Anomaly-detection methods produce zero detection performance and an FNR of 1.0, consistent with their assumption of structurally consistent training and test objects despite PCB-MC’s diverse, unaligned layouts.Missing components may lack strong local visual anomalies and instead require reasoning about expected presence at specific locations.Source: §V · pp. 4–5

Category Level Analysis

• Across most categories, over 85% of missing components are assigned to background, reaching 100% for small LED footprints; larger components fare better, but 68.0% are still missed as background.The confusion matrix shows minimal confusion between component categories, indicating that localization is more problematic than classification once a candidate region exists.Source: Figure 2 · p. 4, §V · p. 5

Quantitative Evaluation

• Samples 1–3 fail when no proposal covers the missing footprint, whereas detection succeeds in samples 4–6 with proposal coverage.The analysis also notes that training data typically lacks explicit negative examples of expected-but-absent objects; the two-stage method partially addresses this by decoupling localization and verification.Source: §V · p. 5, Figure 3 · p. 5

Discussion

• PCB inspection spans appearance detection in task C to structure-aware reasoning under absence in task M, and limitations also interact in the combined task A.Detectors perform well when a task matches their assumptions and degrade when those assumptions do not hold.Source: §V · p. 5

VI. CONCLUSION

• PCB-MC provides a dataset with footprint-level annotations and unified evaluation settings for missing component detection.Source: §VI · p. 6
• Modern detectors have high false negative rates, while anomaly methods fail completely on missing component detection.Source: §VI · p. 6
• Future research will explore structure-aware modeling, including spatial expectations, layout priors, and relational reasoning.The authors expect PCB-MC to serve as a benchmark to stimulate this research in industrial inspection.Source: §VI · p. 6