NCKH – Pilot Scale Computer Vision Non-Invasive
AI-Powered Shrimp Measurement
Measure size · Estimate weight · Shrimp age - Non-invasive
Download CSV CNN YOLOv8 + CNN
Live Camera Feed – Pilot Scale Pond RTSP Ready
Live Camera Feed
Ready for RTSP / USB Camera
Waiting for camera connection
Shrimp #1024 – 4.2cm
Shrimp #1023
Capture Controls
Last measurement: 2 minutes ago Today total: 234 measurements
Latest Results
  • #1024 — 42mm | 8.5g | 45 days Healthy
  • #1023 — 38mm | 7.2g | 40 days Healthy
  • #1022 — 25mm | 3.1g | 25 days Slow
41 mm
7.8 g
42 days
2,456
Size Distribution (mm)
Weight Distribution (g)
Age Distribution (days)
Growth Trend
Size-Weight Correlation
Measurement History - Shrimp Individuals
IDSizeWeightAgeGrowth rateStatusLast measurement
#102442 mm8.5 g45 days+0.9 mm/dayHealthyJust now
#102338 mm7.2 g40 days+0.95 mm/dayHealthy2 min
#102225 mm3.1 g25 days+0.5 mm/daySlow Growth5 min
#102151 mm10.2 g52 days+0.98 mm/dayHealthy8 min
#102032 mm5.8 g30 days+1.07 mm/dayHealthy12 min
92.4%
94.1%
91.8%
1.3 mm
Training Loss vs Validation Loss (CNN) BCE + IoU per epoch
mAP Convergence by Epoch
Precision-Recall Curve (PR Curve)
Size: Predicted vs Actual (mm)
Weight: Predicted vs Actual (g)
Regression MAE theo Epoch (Size & Weight)
CNN Architecture - Shrimp Detection & Measurement
Detection (YOLOv8-n)
  • Backbone: CSPDarknet53
  • Neck: PANet FPN
  • Head: Decoupled (detect + segment)
  • Input: 640×640 | Classes: 1 (shrimp)
  • Batch: 16 | Epochs: 150
Measurement CNN
  • Conv2D(64) → BN → ReLU ×3
  • MaxPool → Conv2D(128) ×2
  • GlobalAvgPool → Dense(256) → Dense(3)
  • Output: [size_mm, weight_g, age_days]
  • Loss: Huber | Optimizer: AdamW
Data Augmentation
  • Random flip, rotation (±15°)
  • Color jitter (brightness, contrast)
  • Mosaic augmentation
  • Train set: 2,400 | val: 600 images
  • Reference marker: 10 mm calibration card