Image Segmentation for AM
Deep learning segmentation for defect detection, microstructure analysis, and in-situ process monitoring
Area
Computer Vision + AM
Papers (250+ cit)
620
Total Citations
380,000+
Architectures
U-Net, Mask R-CNN, DeepLab
Modalities
Optical, CT, SEM, Thermal
mIoU
85-95%
Image segmentation for additive manufacturing uses deep learning to partition images into semantically meaningful regions. In AM, this enables automated defect detection, microstructure quantification, melt pool boundary extraction, and real-time quality monitoring - tasks that previously required expert manual analysis.
Overview
Semantic segmentation assigns class labels to every pixel in an image, enabling quantitative analysis of AM parts and processes. Instance segmentation further distinguishes individual objects (e.g., separate pores), critical for defect characterization.
95%
Defect Detection Acc.
Real-time
Processing Speed
Segmentation Architectures
| Architecture |
Type |
AM Application |
Strength |
| U-Net |
Encoder-Decoder |
Melt pool, microstructure |
Works with small datasets |
| Mask R-CNN |
Instance |
Individual pore detection |
Per-defect statistics |
| DeepLabV3+ |
Semantic |
Layer boundary detection |
Multi-scale features |
| SegFormer |
Transformer |
High-res CT analysis |
Long-range dependencies |
| Attention U-Net |
Attention-gated |
Small defect focus |
Learns where to look |
3D Segmentation
For volumetric data (CT scans), 3D architectures are used:
- 3D U-Net: Extends U-Net to volumetric data
- V-Net: Dice loss for imbalanced datasets
- nnU-Net: Self-configuring for medical/industrial CT
Defect Detection
Defect Types
| Defect |
Imaging |
Segmentation Challenge |
Detection Rate |
| Porosity |
CT, optical |
Small size, irregular shape |
>95% |
| Lack of fusion |
CT |
Elongated, layer-aligned |
>92% |
| Cracks |
SEM, optical |
Thin, branching |
>90% |
| Surface defects |
Optical, profilometry |
Texture variation |
>88% |
| Inclusions |
CT, SEM |
Contrast dependent |
>85% |
Data Augmentation
Critical for AM where labeled defect data is scarce:
- Geometric: Rotation, flipping, elastic deformation
- Intensity: Brightness, contrast, noise addition
- Synthetic defects: GAN-generated or physics-based
- Domain adaptation: Transfer from similar materials
Microstructure Analysis
Segmentation Targets
- Grain boundaries: EBSD data, columnar vs equiaxed
- Phase identification: Alpha/beta in Ti, austenite/martensite
- Precipitates: Carbides in Inconel, intermetallics
- Dendrite structure: Primary/secondary arm spacing
Quantitative Metrics
Segmentation enables automated measurement of:
- Grain size distribution: ASTM E112 equivalent
- Phase fraction: Volume percentage of each phase
- Porosity: Total porosity, size distribution, morphology
- Texture: Preferred orientation from segmented grains
Melt Pool Segmentation
In-Situ Monitoring
Real-time melt pool segmentation from thermal/optical cameras:
- Boundary extraction: Melt pool width, length, area
- Keyhole detection: Critical for avoiding porosity
- Spatter tracking: Ejected particles during printing
- Plume analysis: Vapor/plasma characteristics
Challenges
- High speed: 10,000+ fps required for real-time
- Lighting variation: Intense emission from melt pool
- Motion blur: Fast scan speeds
- Class imbalance: Small melt pool vs large background
Lightweight Models
For real-time deployment:
- MobileNet encoder: Reduced parameters
- EfficientNet: Optimized accuracy/speed
- Knowledge distillation: Small student from large teacher
- TensorRT/ONNX: Hardware acceleration
Key Papers
U-Net: Convolutional Networks for Biomedical Image Segmentation
Ronneberger et al., 2015 | Foundation architecture | 65,000+ citations
Deep learning for defect detection in additive manufacturing
Survey of DL methods for AM quality control | 800+ citations
Automated porosity segmentation in CT images of metal AM parts
U-Net for AM defect detection | 400+ citations
Real-time melt pool monitoring using semantic segmentation
In-situ process monitoring | 300+ citations
View all image segmentation papers →