25-26J-452 - Research-Project

🧵 FabricVision - AI-Driven Fabric Defect Detection & Monitoring System for the Apparel Industry in SRI LANKA

FabricVision is an AI-driven real-time fabric inspection system integrating camera-based acquisition, edge-level enhancement, ML defect detection, and a centralized QC dashboard. It automates inspection, improves defect accuracy, and provides real-time monitoring, traceability, and decision support.

Team Members:

• Member 1 - Nagahawaththa J.C.D - IT22573896 (Team Lead)
• Member 2 - Kamburugamuwa K.S.D - IT22574572
• Member 3 - Jayasinghe J.A.D.T.S - IT22028464
• Member 4 - Kavinda S.G.D - IT22603654

---

System Components:

Component 1 – Intelligent Fabric Capture, Anomaly Pre-Screening & Quality Analytics

Component 1 focuses on capturing only meaningful fabric frames, pre-screening unusual frames, and deciding whether frames are suitable for defect detection. This ensures the downstream defect detection modules receive high-quality and relevant frames, improving overall reliability.

1) Component 1 Workflow (3 Novelty Stages)

Stage 1: Motion-Aware Event-Based Frame Extraction (Novelty 1)

Goal: Reduce redundant frame capture and only extract frames when fabric is moving meaningfully.

How it works:

• Live camera feed/video frames are continuously observed.
• A MobileNetV2-based motion classifier detects fabric presence and motion state (e.g., idle / active).
• Frames are captured only when motion is valid (event-based trigger).

Outputs:

• Selected frames (active motion)
• Logs/metrics: total frames, saved frames, ignored frames, motion state timeline

Why it matters:

• Reduces processing cost and latency
• Prevents sending irrelevant/duplicate frames downstream

Stage 2: Unsupervised Anomaly Pre-Screening (Novelty 2)

Goal: Identify unusual frames early without needing defect labels.

How it works:

• An Autoencoder model (AnomalyAutoencoder128) is trained using only normal fabric frames.
• For each incoming frame, the model calculates reconstruction error (MSE).
• Based on error thresholds, frames are categorized as:
  1. Normal
  2. Borderline
  3. Irregular

Outputs:

• Frame category (normal/borderline/irregular)
• Frame Irregularity Score (FIS) + threshold used
• Only borderline + irregular are forwarded to the next stage

Why it matters:

• Early filtering reduces noise
• Highlights suspicious frames for deeper quality validation

Stage 3: Frame Quality Intelligence & Routing (Novelty 3)

Goal: Decide whether suspicious frames are usable for defect detection or must be enhanced/held.

How it works:

• Borderline/Irregular frames are evaluated via a quality assessment API (rule-based logic for now).

Inputs include:

1. Motion state + confidence
2. Frame type (borderline/irregular)
3. FIS + threshold

The system returns:

1. Frame quality: good / poor
2. Risk level: low / high / critical
3. Action: continue / alert_operator

Routing decisions:

• Usable → Forward to defect detection (or next component)
• Needs enhancement → Forward to Fog enhancement
• Rejected/Hold → Drop/Hold + operator alert + roll-level recommendation

Roll-level decision support (important for panel Q/A):

• System tracks rejected ratio (% poor/held frames).

• If rejected ratio exceeds a threshold → recommend operator actions such as:

  1. check lighting/camera vibration
  2. slow machine speed
  3. re-capture affected segment / re-run inspection

Why it matters:

• Prevents misleading defect detection from blurry/unstable frames
• Protects overall system accuracy by ensuring input quality

---

Component 2 – Title

Add Details about Component 2 here

---

Component 3 – Title

Add Details about Component 3 here

---

Component 4 – Title

Add Details about Component 4 here

---

Technology Stack:

• Frontend: Next.js, Tailwind CSS, Chart.js, Framer Motion
• Backend: FastAPI, Python, OpenCV, TensorFlow/PyTorch
• Machine Learning: MobileNetV2, Autoencoder128, YOLOv9, XGBoost, LSTM

Project Running Script :

Backend Setup:

cd fabapi
pip install -r requirements.txt
uvicorn app:socket_app --host 0.0.0.0 --port 8000 --reload

Frontend Setup:

cd fabricvision-frontend
pnpm install
pnpm run dev

If pnpm is not installed, install it via npm:

npm install -g pnpm