metadata
license: mit
tags:
- image-classification
- computer-vision
- defect-detection
- automotive
- pytorch
- timm
- efficientnet
language:
- ru
pipeline_tag: image-classification
Paint Defect Detector
A binary image classifier that detects paint defects on car body panels using transfer learning with EfficientNetV2-S backbone (via imm).
Model Architecture
- Backbone: EfficientNetV2-S (pretrained, from imm)
- Head: Dropout → Linear(feat_dim, 256) → GELU → Dropout → Linear(256, 2)
- Task: Binary classification — clean vs defect
Training
- Optimizer: AdamW with cosine annealing LR scheduler
- Loss: CrossEntropyLoss with label smoothing
- Augmentations: Albumentations pipeline
- Metrics: AUC-ROC, F1, Accuracy
Inference
The project includes a FastAPI REST API (src/api.py) for serving predictions, and a Grad-CAM visualisation layer for model explainability.
Project Structure
src/ config.py # Hyperparameters and paths dataset.py # Dataset and data loaders model.py # DefectClassifier model train.py # Training loop infer.py # Inference utilities api.py # FastAPI inference server prepare_data.py # Data preparation script requirements.txt
Requirements
See equirements.txt. Key dependencies: orch, imm, lbumentations, astapi, grad-cam.