How wangerniuniu/fabricdefectdetection works
Project type and purpose
• This is a research‐oriented deep learning project (“FabricDefectDetection”) that implements a real‐time fabric defect segmentation network in PyTorch.
• Its primary function is loading labeled fabric defect images, passing them through a convolutional segmentation model (defined in model.py), and producing defect masks.
File structure
• Top level:
• HKSEG/… : six subfolders (BrokenEnd, Hole, Knots, NettingMultiple, ThickBar, ThinBar), each containing image datasets and any annotation attachments.
• model.py : the core model definition, training and inference scripts (often combined).
• README.md : project overview, prerequisites, model visuals and performance stats.
README for additional insights
• Confirms use of Ubuntu + Python3.5 + PyTorch1.0 + CUDA9.0
• Mentions external labeling tool for annotation.
Main components
a. Data Component
• Raw image datasets (HKSEG/… folders)
b. Data Pipeline
• ImageReader & Preprocessing (crop, resize, normalization)
• DataLoader (batching, shuffling)
c. Model Component
• Convolutional Neural Network (encoder‐decoder segmentation architecture from image/net.png)
• Intermediate blocks (residual/dense blocks as per image/details.png)
d. Training Loop
• Loss computation (e.g., cross‐entropy, IoU loss)
• Optimizer (SGD/Adam), learning‐rate scheduler
• Backpropagation on GPU (CUDA 9.0)
e. Inference & Postprocessing
• Softmax threshold, morphological cleaning
• Overlay masks on original images
f. Outputs & Evaluation
• Visual results (result1.png, result2.png)
• FLOPs & model‐size metrics (flops.jpg, modelsize.jpg)
g. External Services
Relationships and interactions
• Data → Preprocessing → DataLoader → Model → Loss & Optimizer ↺ Model (training)
• For inference: DataLoader → Model → Postprocessing → Result Storage/Visualization
• The GPU runtime (CUDA) underpins Model and Training Loop operations.
Architectural patterns & design principles
• Modular pipeline: separation of data ingestion, model definition, and training/inference logic.
• Encoder‐decoder segmented network (common segmentation pattern).
• Hardware abstraction: PyTorch’s CUDA backend abstracts GPU details.