How vidhigadge/steelvision-dl works
SteelVision is a steel-defect research prototype with a user-facing inspection workflow and supporting model-development experiments. In standard analysis, a user submits an image through the interface; the API classifies it, reports confidence and entropy, and can request Grad-CAM and localization with mapped defect boxes. A separate localization experiment varies input resolution. The graph also preserves dataset, training, super-resolution, reliability, and deployment-optimization capabilities documented in the README. Source excerpts establish API-to-service and training-to-model calls; unsampled UI wiring and several research links are shown only as unverified where appropriate.
Main entry points: resnet50_finetune.py Fine-Tuned ResNet · training Model Training · main.py FastAPI Endpoints · evaluation Model Evaluation
Components 20
User Interface
- Streamlit Interface
src/ui/app.py
Inspection Services
- FastAPI Endpoints
src/api/main.py - ONNX Classifier
src/api/main.py - Grad-CAM Service
src/explainability/gradcam_service.py - YOLO Localization
src/detection/localization_service.py - Box Coordinate Mapping
src/detection/coordinate_mapping.py
Model Research
- Baseline CNN
src/models/cnn.py - ResNet-50
src/models/resnet50.py - Fine-Tuned ResNet
src/models/resnet50_finetune.py - Model Training
src/training - Model Evaluation
src/evaluation - Resolution Experiments
src/evaluation/evaluate_resolution.py - Super-Resolution Study
src/super_resolution - Uncertainty and Calibration
src/reliability - SHAP Explanations
src/explainability/shap_explain.py
Dataset Workflow
- Image Splits
data/splits - Dataset Preparation
src/data - Classification Loaders
src/data/loaders.py
Deployment Optimization
- ONNX Optimization
src/optimization
Other components
- Inspector
Connections 18
- Inspector → Streamlit Interface: submits image
- Streamlit Interface → FastAPI Endpoints: requests analysis
- FastAPI Endpoints → ONNX Classifier: runs inference
src/api/main.py - FastAPI Endpoints → Grad-CAM Service: requests explanation
src/api/main.py - FastAPI Endpoints → YOLO Localization: requests localization
src/api/main.py - YOLO Localization → Box Coordinate Mapping: maps boxes
src/detection/localization_service.py - Model Training → Classification Loaders: loads batches
src/training/train_resnet_finetune.py - Model Training → Fine-Tuned ResNet: trains model
src/training/train_resnet_finetune.py - Model Training → Baseline CNN: trains baseline
src/training/train_cnn.py - Model Training → ResNet-50: trains transfer model
src/training/train_resnet.py - Model Evaluation → Classification Loaders: loads evaluation data
src/evaluation/evaluate_resnet_finetune.py - Model Evaluation → Fine-Tuned ResNet: evaluates model
src/evaluation/evaluate_resnet_finetune.py - Resolution Experiments → Fine-Tuned ResNet: measures accuracy
src/evaluation/evaluate_resolution.py - Super-Resolution Study → Fine-Tuned ResNet: compares classification
README.md - Uncertainty and Calibration → Fine-Tuned ResNet: analyzes predictions
src/reliability/uncertainty.py - SHAP Explanations → Fine-Tuned ResNet: explains predictions
src/explainability/shap_explain.py - ONNX Optimization → Fine-Tuned ResNet: optimizes inference
README.md - Dataset Preparation → Image Splits: prepares splits