[ML/Backend] Integrate MobileNetV3 Species Classifier into /api/v1/scan Endpoint

[jpdevhub]

Problem

The MobileNetV3-Small species model (Models/species_mobilenetv3.pth) was merged in but is not yet wired into the FastAPI scan pipeline. The /api/v1/scan endpoint currently returns freshness data only; it does not run species classification.

Goal

Connect the pre-trained .pth model to the scan endpoint so the API response includes species classification alongside freshness scores.

Acceptance Criteria

• backend/species_classifier.py created — loads Models/species_mobilenetv3.pth at startup (not on every request) using torch.load with map_location='cpu'
• /api/v1/scan response extended with a new species object:

  {
    "species": {
      "name": "Rohu Carp",
      "confidence": 0.945,
      "top3": [
        { "name": "Rohu Carp", "confidence": 0.945 },
        { "name": "Catla Carp", "confidence": 0.038 },
        { "name": "Mrigal Carp", "confidence": 0.017 }
      ]
    }
  }

• Model inference runs with torch.no_grad() for performance
• Image pre-processing matches the training pipeline: resize to 224×224, ImageNet normalization
• If species model fails for any reason, the scan still succeeds (species is null, not a 500 error)
• Unit test added: backend/tests/test_species_classifier.py

Supported Species (10 classes)

Rohu Carp, Catla Carp, Mrigal Carp, Pangas, Basa, Tilapia, Pomfret, Kingfish, Mackerel, Sardine

Notes

• Model is already in Models/species_mobilenetv3.pth (6.0 MB, MobileNetV3-Small, 94.5% val accuracy)
• Do NOT reload the model on every request — load once at module import time using a module-level variable
• This task is a prerequisite for the frontend UI issue (species display on ResultsPage)

[arcgod-design]

Hi @jpdevhub! I'd love to work on this as part of SSoC 2026. I'm @arcgod-design.

My Approach:

• Create backend/species_classifier.py that loads Models/species_mobilenetv3.pth at startup using torch.load with map_location='cpu'
• Run inference on the uploaded image and return {name, confidence, top3} object
• Extend the /api/v1/scan response to include the species field
• Ensure inference runs in parallel with or after freshness analysis without blocking
• Keep startup load-once pattern to avoid reloading on every request
• Handle model-not-found gracefully (return species: null if weights not present)

Could you please assign this issue to me? Ready to start immediately!

[jpdevhub]

go ahead