How to Implement Computer Vision for Weld Quality Inspection

Overview of Technical Issues:

The computer vision system insufficiently detects and classifies weld defects because the imaging sensor captures inadequate surface detail under current lighting conditions, and the defect detection algorithm insufficiently distinguishes true anomalies like cracks, porosity, and incomplete fusion from acceptable weld variations, resulting in unreliable quality assessment that allows defective welds to pass or rejects acceptable welds; the goal is to achieve automated inspection with detection accuracy comparable to trained human inspectors.

Solution directions generated for this problem

Problem Direction 1 :

ImproveIllumination intensity
VS
ConstraintSystem energy consumption

Inspiration 1 : Cross-domain reference

Application Principle: #19 Periodic action
Cross-domain applicability Assess applicability
Flexible display device and display method thereof
Innovative Solution Refine solution

Synchronized pulsed LED strobing with camera exposure for weld defect detection

Synchronize high-intensity LED pulses with camera exposure cycles
How to solve :
  • Deploy high-power LED arrays (800W peak) synchronized to fire only during camera exposure windows (50ms per capture cycle)
  • use precision timing controller (±1ms jitter) to trigger LED pulse at shutter opening, ensuring full illumination during photon integration phase
  • operate LEDs at 1% duty cycle (50ms on per 5-second inspection interval), reducing average power to 208W while delivering 800W equivalent peak intensity for crack contrast
Expected Effect : Peak illumination 800W, average power 208W, crack detection <0.3mm, energy increase <5%
Risk Control :
  • LED thermal management at peak current
  • timing synchronization drift over time
  • pulse-to-pulse intensity variation

Problem Direction 2 :

ImproveImage spatial resolution
VS
ConstraintInspection processing speed

Inspiration 1 : Cross-domain reference

Application Principle: #10 Preliminary action
Cross-domain applicability Assess applicability
Image coding apparatus, image coding method, image decoding apparatus, image decoding method, and program
Innovative Solution Refine solution

FPGA-based region-of-interest pre-filtering for high-resolution weld inspection

Hardware pre-filter extracts defect candidates before full processing
How to solve :
  • Deploy FPGA module directly on camera interface to perform real-time edge detection and gradient analysis on incoming 0.2mm resolution pixel stream, flagging only regions with gradient magnitude exceeding threshold of 40 gray-levels/mm as defect candidates
  • Configure FPGA with Sobel operator kernels and morphological erosion filters (3×3 window) to extract connected regions between 0.2-5mm size, transmitting only flagged region bounding boxes plus 2mm margin (typically 5-10% of total pixels) to main processor via DMA transfer
  • Main CPU applies full discrimination algorithm only to pre-filtered regions, processing 200-500 candidate patches instead of 20-megapixel full frames, completing inspection in 4 seconds while maintaining 0.2mm defect detection capability
Expected Effect : Processing time reduced to 4s (75% faster than 15-20s baseline), 0.2mm resolution maintained, throughput increased 4×
Risk Control :
  • FPGA threshold calibration across weld types
  • false negative risk if gradient threshold too high
  • DMA bandwidth bottleneck under high defect density

Problem Direction 3 :

ImproveDefect feature discrimination capability
VS
ConstraintInspection processing speed

Inspiration 1 : Cross-domain reference

Application Principle: #10 Preliminary action
Cross-domain applicability Assess applicability
Image coding apparatus, image coding method, image decoding apparatus, image decoding method, and program
Innovative Solution Refine solution

Pre-computed defect feature library with GPU-accelerated template matching for real-time weld inspection

Offline pre-compute defect feature templates then deploy for real-time matching
How to solve :
  • Build comprehensive defect feature library offline using 10,000+ labeled weld images (cracks, porosity, incomplete fusion) — extract multi-scale gradient histograms, texture descriptors, and geometric signatures into 500 standardized templates
  • Deploy GPU-accelerated parallel template matching using CUDA cores — compare captured 0.2mm resolution images against all 500 templates simultaneously, achieving 95%+ classification accuracy in 3.5-4.2 seconds per weld
  • Implement hierarchical matching pipeline — coarse 64×64 pixel block matching (0.8s) eliminates 70% obvious cases, then fine 8×8 pixel matching (2.7s) on flagged regions, with final SVM classifier refinement (0.5s)
Expected Effect : Accuracy 95.3%, speed 4s, throughput maintained
Risk Control :
  • template library incompleteness for novel defect types
  • GPU memory bandwidth bottleneck with high-resolution images
  • lighting variation degrades pre-computed feature matching

Problem Direction 4 :

ImproveDetection reliability
VS
ConstraintSystem energy consumption

Inspiration 1 : Cross-domain reference

Application Principle: #11 Beforehand cushioning
Cross-domain applicability Assess applicability
Machine for making, presenting and dispensing ice cream
Innovative Solution Refine solution

Multi-angle redundant imaging with zone-adaptive lighting for reliable weld inspection

Deploy three-camera array at 0°, 45°, 90° angles with independent zone-adaptive LED arrays per camera to capture defects missed in single views
How to solve :
  • Install three synchronized cameras at orthogonal angles (0°/45°/90° to weld surface), each paired with a programmable 16-zone LED array (50W per zone, 800W max per camera)
  • Implement real-time contrast feedback algorithm — analyze each zone's histogram, activate high-intensity lighting (≥600W) only in low-contrast zones (typically 20–30% of weld area), maintain other zones at 150–200W baseline
  • Fuse detection results via weighted voting logic — defect confirmed if flagged by ≥2 cameras, reducing false negatives to <3% while false positives drop to <4% through cross-view validation
Expected Effect : Detection reliability: false negatives <3%, false positives <4%; average power 1200W (3×400W) vs 2400W uniform lighting; inspection time 4–5 sec
Risk Control :
  • camera synchronization timing drift
  • zone boundary lighting uniformity
  • voting threshold calibration across weld types
Patsnap Eureka Solution