Scoring Pipeline

The image-scoring project uses a multi-stage pipeline to assess image quality using neural network models. This skill covers the architecture, key modules, and how to safely make changes.

Architecture Overview

Input (folder/file)
  → PrepWorker (RAW conversion, thumbnails, skip checks)
    → ScoringWorker (GPU inference: MUSIQ, LIQE, TOPIQ)
      → ResultWorker (DB upsert, XMP metadata write, cleanup)

The pipeline uses producer-consumer queues with threading.Thread workers.

Key Modules

Module	Purpose
modules/pipeline.py	Worker classes: PrepWorker, ScoringWorker, ResultWorker, ImageJob dataclass
modules/scoring.py	ScoringRunner — orchestrates batch/single scoring, Fix DB, and metadata repair
modules/engine.py	BatchImageProcessor — low-level batch processing engine
modules/config.py	load_config(), get_config_value() — reads config.json
modules/utils.py	Path conversion (WSL ↔ Windows), hashing, thumbnail utilities

Model Wrappers

File	Model	Framework	Score Range
musiq/run_all_musiq_models.py	MUSIQ (SPAQ, AVA, KonIQ, PaQ2PiQ)	TensorFlow	0–100 (normalized to 0–1)
modules/liqe.py	LIQE (language-image quality evaluator)	PyTorch (pyiqa)	1–5 (normalized to 0–1)
modules/topiq.py	TOPIQ-IAA	PyTorch (pyiqa)	0–1
modules/qalign.py	Q-Align	PyTorch	1–5

Scoring Formulas (Current)

General  = 0.50 × LIQE + 0.30 × AVA + 0.20 × SPAQ
Technical = LIQE (primary)
Aesthetic = 0.60 × AVA + 0.40 × SPAQ

All composite scores are stored as floats 0.0–1.0 in the database.

Hybrid Environment

• Windows: Runs the WebUI, PostgreSQL Docker, and file management.
• WSL 2 (Linux): Runs GPU inference (TensorFlow + PyTorch with CUDA).
• CRITICAL: Database access from WSL must use TCP (port 5432). PostgreSQL runs in Docker on the Windows host.

How to Add a New Model

1. Create a wrapper in modules/ (e.g., modules/newmodel.py) that exposes a score(image_path) → float method.
2. Import and call it in ScoringWorker.process() inside modules/pipeline.py.
3. Add the raw score column to the DB schema in modules/db.py → _init_db_impl().
4. Update the composite formulas in modules/scoring.py → ScoringRunner.fix_image_metadata().
5. Add normalization logic if the model's native range isn't 0–1.

Configuration

All config lives in config.json at the project root. Key sections:

• database — filename, user, password
• scoring — force_rescore_default, model weights