YOLOv8 · EfficientNetV2S · Grad-CAM · U²-Net

Freshly
Fishy

AI-Driven Marine Quality Analysis. Two-stage computer vision detects, segments, classifies, and explains fish freshness in under five seconds.

Scroll

Two-Stage Pipeline

Detect. Segment.
Understand.

YOLOv8 isolates the fish with a sub-8 ms bounding box. U²-Net traces the exact body outline pixel by pixel. EfficientNetV2S reads freshness from eyes, gills, and skin texture. Grad-CAM makes the decision explainable.

YOLOv8 detection

7.4 ms · 95.6% precision

U²-Net segmentation

Pixel-accurate fish mask

EfficientNetV2S

22k-sample fine-tuned CNN

Grad-CAM heatmap

Background-suppressed XAI

YOLO Detector v8 · Live Feed
REC
fish 0.97
FRESH · 91.2%
✓ Auto Approved
7.4 ms / frame

Model performance

Numbers that matter

0+

Training images

Fresh · Not Fresh balanced

0.0%

Detection precision

YOLOv8 on held-out test set

0.0 ms

Per-frame inference

135+ FPS real-time throughput

0.0%

mAP@50 score

Industry-grade detection quality

0.000%

Classifier test accuracy

EfficientNetV2S on held-out test set

mAP@50 · 0.938Precision · 0.956Recall · 0.875mAP@50-95 · 0.757Test Accuracy · 94.38%AUC · 0.982F1 · 0.914Inference · 7.4 msDataset · 22,000 imgsClasses · 2Input · 224×224Backbone · EfficientNetV2SDetector · YOLOv8mmAP@50 · 0.938Precision · 0.956Recall · 0.875mAP@50-95 · 0.757Test Accuracy · 94.38%AUC · 0.982F1 · 0.914Inference · 7.4 msDataset · 22,000 imgsClasses · 2Input · 224×224Backbone · EfficientNetV2SDetector · YOLOv8m

Pipeline

Five stages.
One verdict.

1

Detection

YOLOv8m

High-res 1280px pass isolates the fish with a tight bounding box in under 8 ms.

2

Segmentation

U²-Net / rembg

Nested U-Net traces the fish silhouette pixel-by-pixel, producing a clean foreground mask.

3

Enhancement

CLAHE + BG Blur

Contrast-limited adaptive histogram equalisation sharpens the fish; background is blurred, not zeroed.

4

Classification

EfficientNetV2S

Fine-tuned backbone + GELU head delivers [P(fresh), P(not_fresh)] with a calibrated confidence gate.

5

Explanation

Grad-CAM

Gradients w.r.t. the last Conv2D, masked by the fish silhouette, highlight eyes · gills · skin.

Explainability preview

See what the AI sees

Hover over the fish below to reveal the Grad-CAM attention heatmap. Warm regions drove the freshness decision.

Hover to reveal heatmap
High attention
Medium
Low

Live inference

Try it on your own fish

Upload a photo or use your camera. The full pipeline runs locally — detection, segmentation, classification, and Grad-CAM — and returns in under five seconds.

🐟FreshlyFishyAI Freshness Analysis
🐠

No analysis yet

Upload a fish photo or use your camera — results appear here instantly.