How to Select Computer Vision Architectures for Underwater Welding Inspection

Overview of Technical Issues:

The vision processing architecture insufficiently extracts defect features from underwater welding images degraded by turbidity and light scattering, leading to unreliable defect detection accuracy; the goal is to select an architecture capable of robust feature extraction under harsh underwater conditions to enable reliable automated welding quality inspection.

Solution directions generated for this problem

Problem Direction 1 :

ImproveFeature extraction robustness
VS
ConstraintProcessing computational complexity

Inspiration 1 : Cross-domain reference

Application Principle: #1 Segmentation
Cross-domain applicability Assess applicability
Method of transmitting and receiving broadcasting channel and device therefor
Innovative Solution Refine solution

Modular turbidity-adaptive defect detection pipeline with specialized processing branches

Divide pipeline into specialized branches for different turbidity levels
How to solve :
  • Deploy a lightweight turbidity classifier (3-layer CNN, 0.08 GFLOPs, 15ms latency) as first stage to measure image SNR and route to specialized branches
  • For clear water (SNR ≥18dB), use fast shallow detector (2-layer CNN, 0.4 GFLOPs) targeting high-contrast defects
  • For turbid conditions (SNR <18dB), activate targeted denoising-detection module (5-layer ResNet with spatial attention on crack regions only, 1.2 GFLOPs) processing enhanced features
Expected Effect : Average computation 0.65 GFLOPs (35% reduction vs. universal robust architecture); accuracy 82% clear water, 78% turbid conditions; processing time 45ms per frame
Risk Control :
  • turbidity classification error causing misrouting
  • branch transition instability at SNR threshold boundaries
  • insufficient training samples for turbid-condition branch

Problem Direction 2 :

ImproveDefect signal discriminability
VS
ConstraintArchitecture training data requirement

Inspiration 1 : Cross-domain reference

Application Principle: #2 Taking out
Cross-domain applicability Assess applicability
Dynamic spatial scene analysis
Innovative Solution Refine solution

Physics-guided feature extraction for data-efficient defect detection

Extract physics-based invariant features before learning
How to solve :
  • Apply underwater image formation model (I=J·t+A·(1-t)) to decompose each image into direct transmission component J and backscatter A, isolating defect signals from turbidity noise
  • Extract degradation-invariant edge features using Sobel gradients on restored J component at 0°/45°/90°/135° orientations, generating 4-channel edge maps where crack boundaries remain stable across turbidity (contrast variation <15%)
  • Train compact 3-layer CNN classifier (0.5M parameters: Conv64-Conv128-FC256) exclusively on physics-extracted edge features, requiring only 1,800 labeled samples (1,200 real + 600 augmented via rotation/scaling) to achieve 82% detection accuracy across 0.5-5 NTU turbidity range
Expected Effect : Data requirement reduced 82%; crack detection accuracy 82% across turbidity; processing time 45ms per frame
Risk Control :
  • Image formation model parameter calibration errors
  • edge extraction sensitivity to extreme scattering (>5 NTU)
  • insufficient augmentation diversity for rare defect types

Problem Direction 3 :

ImproveDetection accuracy reliability
VS
ConstraintArchitecture training data requirement

Inspiration 1 : Cross-domain reference

Application Principle: #6 Universality
Cross-domain applicability Assess applicability
Failed and censored instances based remaining useful life (RUL) estimation of entities
Innovative Solution Refine solution

Multi-task shared encoder for underwater weld defect detection with limited data

Multi-task shared encoder for data efficiency
How to solve :
  • Design a shared convolutional encoder (ResNet-18 backbone, 11M parameters) trained simultaneously on three related tasks: defect detection (binary classification), turbidity level estimation (regression 0-5 NTU), and weld seam segmentation (pixel-wise labeling)
  • Split 1,200 samples into task-specific subsets: 600 for defect detection, 400 for turbidity estimation, 200 for segmentation—shared encoder learns robust underwater-invariant features from all tasks, while task-specific heads (2-layer MLPs, 50K parameters each) specialize on individual objectives
  • Implement gradient balancing during training: weight task losses as 0.5:0.3:0.2 (defect:turbidity:segmentation) to prioritize primary detection task while leveraging auxiliary tasks for feature robustness—train 80 epochs with Adam optimizer (learning rate 1e-4), batch size 16, data augmentation (horizontal flip, Gaussian noise σ=0.02, brightness ±15%)
Expected Effect : Detection accuracy 78-82% across 0.5-4 NTU turbidity (vs 45-85% baseline); effective training data utilization 3.2× higher; inference time 45ms per frame on NVIDIA Jetson Xavier
Risk Control :
  • task loss imbalance causing encoder bias toward auxiliary tasks
  • insufficient segmentation labels (200 samples) degrading shared features
  • overfitting on limited turbidity range in training set

Problem Direction 4 :

ImproveFeature extraction robustness
VS
ConstraintArchitecture training data requirement

Inspiration 1 : Cross-domain reference

Application Principle: #27 Cheap short-living objects
Cross-domain applicability Assess applicability
Stabilized antibody-containing liquid formulations
Innovative Solution Refine solution

Modular specialist network library with turbidity-triggered routing for robust defect detection

Deploy specialist network library with real-time turbidity routing
How to solve :
  • Build a library of 5 compact specialist CNNs (each 0.8-1.2M parameters), each trained on 200-300 samples from specific turbidity ranges: 0-1 NTU (clear), 1-2.5 NTU (low), 2.5-5 NTU (medium), 5-10 NTU (high), >10 NTU (extreme)
  • Integrate real-time turbidity sensor (optical backscatter probe, 10Hz sampling) at camera housing to measure NTU within ±0.3 accuracy, triggering routing logic to select appropriate specialist network within 15ms
  • Train each specialist on turbidity-specific augmentation: clear-water specialist uses geometric transforms only, turbid specialists add Jaffe-McGlamery scattering simulation with matched NTU parameters, ensuring each network optimizes for its operational envelope
Expected Effect : 82% accuracy across all turbidity levels using 1,500 total samples; 40% faster inference than universal model; modular retraining capability
Risk Control :
  • turbidity sensor calibration drift in long-term deployment
  • routing threshold boundary cases causing specialist mismatch
  • uneven defect type distribution across turbidity ranges during training
Patsnap Eureka Solution