본문 바로가기 주메뉴 바로가기

(주)엠제이비전테크

PCB

Development background

A PCB appearance inspection system whose deep learning model, trained on good and defective images, classifies foreign matter, surface scratches, solder defects and contamination by type, reducing over-detection and eliminating missed defects

Solution

Features

PCB inspection with accuracy boosted by AI deep learning

  • 01

    AI Deep Learning Defect Classification

    The deep learning model learns the shape, texture and position of defects to tell whether each is foreign matter, a scratch, a solder defect or contamination. Non-defect variations such as light reflections or board color differences are learned as normal, reducing over-detection.

    4 defect classes / classification accuracy 99%+ / over-detection under 5%

  • 02

    Foreign Matter · Scratches · Solder · Contamination at Once

    A single inspector detects all four defect types in one capture. Foreign matter and contamination are judged by color and texture, scratches by linear patterns, and solder defects by fillet shape and gloss, so no separate inspection step is needed per defect.

    4 defect types / single capture / minimum detection size 50 μm

  • 03

    Deep Learning and Rule-Based, Each Doing Its Part

    Deep learning handles irregular appearance defects, while rule-based tools handle items with clear criteria such as component position, dimensions and polarity. Combining both keeps defects out of passed products while reducing secondary visual inspection caused by over-detection.

    0% defects in passed products / over-detection ≤5% / 70% less secondary visual inspection

  • 04

    On-Site Training and Quality Data Integration

    When a new defect type appears, images are labeled on site and the model is retrained and applied immediately. Results are stored by board ID, defect position and type, linked to MES quality data, and defect trends can be reviewed as statistics.

    Retraining applied within 1 day / board-level history / MES integration

Function

  • Foreign Matter Inspection

    The deep learning model learns textures that differ from the background, distinguishing foreign matter even when the color is similar

    Targets

    Metal fragmentsSolder ballsFibers · dust

  • Surface Scratch Inspection

    Coaxial and dark-field illumination raise contrast even for shallow scratches, and deep learning tells pattern edges apart from scratches

    Targets

    Resist scratchesExposed copperSilkscreen damage

  • Solder Defect Inspection

    Deep learning judges shape variations that rule-based AOI tends to miss

    Targets

    Excess solder · bridgesLifted leadsInsufficient solderComponent shift · missing · reversed polarity

  • Contamination Detection

    Deep learning learns the color and gloss distribution of a normal board surface to judge subtle differences

    Targets

    Flux residueFingerprints · oilDiscoloration · oxidationStains

Detailed Functions

  • Automatic Defect Classification

    Deep learning distinguishes foreign matter, scratches, solder defects and contamination

  • Defect Location Map

    Shows defect positions and types on the board layout

  • Dimension · Position · Polarity

    Rule-based measurement of component shift, polarity and pad dimensions

  • Board ID Recognition

    Manages per-board history via barcode, 2D code and OCR

  • Defect Statistics · SPC

    Occurrence rate by type, process trends, control charts

  • MES · Equipment Integration

    Sends judgment results, integrates with rejectors and PLCs

Inspection Items

Inspection Items
Defect Type Inspection Area Defect Details Lighting · Optics Judgment Method
Foreign matter Board surface, between components Metal fragments, solder balls, fibers/dust, flux lumps Ring + coaxial Deep learning + size measurement
Surface scratches Resist, copper, silkscreen Scratches, exposed copper, silkscreen damage Coaxial + dark-field Deep learning + length/width measurement
Solder defects SMT mounting areas Insufficient, cold joints, bridges, solder balls, lifted leads, shift/missing/polarity Multi-angle ring (RGB) Deep learning shape classification + rule-based position/polarity
Contamination Entire board, connectors/pads Flux residue, fingerprints/oil, stains, discoloration Dome + coaxial Deep learning + area measurement
상단으로