AI Detection of Fish Diseases from Video and Photos

AI Detection of Fish Diseases from Video and Photos

An ichthyopathologist can inspect 200–500 fish per day. Our CV-based fish disease detection system processes 10,000+ individuals per hour and logs every detected sign tied to a fish ID. We build such turnkey systems for fish farms in aquaculture. Our stack: YOLOv8, EfficientNet, PyTorch. Over 5 years, we have deployed solutions for 12 fish species, including Atlantic salmon, trout, and tilapia. Average customer savings amount to 2.5 million RUB per year due to reduced mortality.

Why a CV System Outperforms an Ichthyopathologist

Humans get tired and miss up to 30% of sick fish under heavy workload. A CV model works 24/7 without loss of concentration. Moreover, the system records exact coordinates of each defect, allowing health dynamics tracking. We confirm this in practice: after deploying our system, clients reduce mortality by 15–20% through early outbreak detection.

Diseases and Their Visual Signs

Each disease has specific visual markers:

Detector Architecture

Two-stage pipeline:

Stage 1: Fish detection — YOLOv8-seg isolates each individual in the frame. This is necessary to separate overlapping fish and normalize crops before classification.

Stage 2: Pathology classification — a multi-label classifier runs on each fish crop. Multi-label is important: a single fish may have both ulcers and exophthalmia.

from ultralytics import YOLO
import torch
from torchvision import models, transforms

# Fish detector
fish_detector = YOLO('fish_detector_yolov8m.pt')

# Pathology classifier (multi-label)
disease_classifier = models.efficientnet_b3(pretrained=False)
disease_classifier.classifier[1] = torch.nn.Linear(
    disease_classifier.classifier[1].in_features,
    num_diseases  # e.g., 12 disease classes
)

def analyze_fish(frame):
    # Fish detection
    detections = fish_detector(frame, conf=0.5)[0]
    results = []
    for box in detections.boxes:
        crop = frame[int(box.xyxy[0,1]):int(box.xyxy[0,3]),
                     int(box.xyxy[0,0]):int(box.xyxy[0,2])]
        # Pathology classification
        tensor = preprocess(crop).unsqueeze(0).cuda()
        with torch.no_grad():
            logits = disease_classifier(tensor)
            diseases = torch.sigmoid(logits) > 0.5  # multi-label
        results.append({'box': box.xyxy[0], 'diseases': diseases})
    return results

How to Calibrate the Model for a Specific Fish Species?

Each species has its own normal pigmentation and anatomy. Atlantic salmon, trout, tilapia — different baselines. We use fine-tuning or separate classifier heads for each species. Fish size also matters: fry (5–10 g) and market-size fish (2–5 kg) require different perspectives and detail resolutions. We add size normalization via estimated body length.

Gills: A Separate Challenge

Gill pathologies are not visible externally without special inspection. For primary processing lines: automated capture of the gill cover image in open position + a specialized classifier. Gill condition classification by color and texture: healthy (dark red, firm) / pale (anemia, possibly oxygen deprivation) / whitened (bacterial gill disease) / brown (methemoglobinemia from nitrites). CNN on gill crops: accuracy 0.88 across four classes.

Comparison of Manual Inspection and CV System

Implementation Process

1. Production analysis: study the conveyor, lighting, fish species (1–2 weeks).
2. Data collection: record video from customer cameras, annotate with ichthyopathologist (2–4 weeks).
3. Model development: train detector and classifier on your data (3–6 weeks).
4. Integration: install software on server, connect to SCADA (1–2 weeks).
5. Testing: validation run on 10,000+ individuals, threshold tuning (1–2 weeks).
6. Staff training: 2–3 days on site, documentation.

What’s Included

• Ready-to-use detection and classification model (Docker image).
• API for integration (REST/gRPC).
• Web interface for viewing results and exporting reports.
• Operation and calibration documentation.
• Model warranty — 6 months of free updates when the species mix changes.

We assess your project in 2 days. Contact us for a preliminary analysis — send us video from your conveyor, and we will return a prototype detection result.

Timelines

Basic detection system (video stream, 5–8 diseases): 8–12 weeks. Extended module with gill analysis and multiple fish species: 14–20 weeks. Cost is calculated individually.

Investment in the system pays off in 6–12 months.