src/data/nist302_loader.py

from __future__ import annotations

from dataclasses import dataclass
from pathlib import Path
from typing import Iterable

import torch
from PIL import Image
from torchvision import transforms

from .base import FingerprintSample


IMAGE_EXTS = {".png", ".jpg", ".jpeg", ".bmp", ".tif", ".tiff"}


@dataclass
class NIST302Paths:
    """Root folders for three NIST SD302 subsets."""

    root_302a: str
    root_302b: str
    root_302d: str


class NIST302Loader:
    """Unified loader for NIST SD302 A/B/D variants.

    The loader parses metadata from path and filename patterns currently present
    under the workspace dataset tree.
    """

    def __init__(self, image_size: int = 224):
        self.transform = transforms.Compose(
            [
                transforms.Resize((image_size, image_size)),
                transforms.ToTensor(),
            ]
        )

    def discover(self, paths: NIST302Paths) -> list[dict[str, str]]:
        records: list[dict[str, str]] = []
        for root_str, dataset_name in [
            (paths.root_302a, "nist_sd302a"),
            (paths.root_302b, "nist_sd302b"),
            (paths.root_302d, "nist_sd302d"),
        ]:
            if not root_str:  # skip empty roots — Path("") resolves to cwd and scans everything
                continue
            records.extend(self._scan_subset(Path(root_str), dataset_name))
        return records

    def iter_samples(
        self, records: Iterable[dict[str, str]]
    ) -> Iterable[FingerprintSample]:
        for rec in records:
            image = Image.open(rec["image_path"]).convert("L")
            tensor = self.transform(image)
            yield {
                "image": tensor,
                "identity_id": rec["identity_id"],
                "finger_id": rec["finger_id"],
                "sensor_id": rec["sensor_id"],
                "dataset": rec["dataset"],
                "image_path": rec["image_path"],
            }

    def _scan_subset(self, subset_root: Path, dataset_name: str) -> list[dict[str, str]]:
        if not subset_root.exists():
            return []

        records: list[dict[str, str]] = []
        for path in sorted(subset_root.rglob("*")):
            if not path.is_file() or path.suffix.lower() not in IMAGE_EXTS:
                continue
            meta = self._parse_metadata(path, subset_root, dataset_name)
            if meta is not None:
                records.append(meta)
        return records

    @staticmethod
    def _parse_metadata(
        file_path: Path,
        subset_root: Path,
        dataset_name: str,
    ) -> dict[str, str] | None:
        rel_parts = file_path.relative_to(subset_root).parts
        stem_tokens = file_path.stem.split("_")
        # SD302-A filenames have 4 tokens: {subject}_{sensor}_{captype}_{finger}
        # SD302-B/D filenames have 5 tokens: {subject}_{sensor}_{dpi}_{captype}_{finger}
        if len(stem_tokens) < 4:
            return None

        identity_id = stem_tokens[0]
        # NIST SD302 identity IDs are 8-digit numbers; reject non-fingerprint files early
        if not (identity_id.isdigit() and len(identity_id) == 8):
            return None

        # Heuristic finger id from filename tail. This should be audited before
        # strict cross-sensor experiments, matching the plan's risk note.
        finger_token = stem_tokens[-1]
        finger_id = f"F{int(finger_token):02d}" if finger_token.isdigit() else finger_token

        # Sensor token is stabilized from the first 1-3 directory markers that
        # capture device/capture style differences.
        sensor_tokens = rel_parts[:-1]
        sensor_id = "_".join(sensor_tokens[:3]) if sensor_tokens else "unknown"

        return {
            "identity_id": identity_id,
            "finger_id": finger_id,
            "sensor_id": sensor_id,
            "dataset": dataset_name,
            "image_path": str(file_path),
        }


def to_batch(samples: list[FingerprintSample]) -> dict[str, torch.Tensor | list[str]]:
    """Collate helper for lists emitted by NIST302Loader.iter_samples."""

    images = torch.stack([s["image"] for s in samples], dim=0)
    return {
        "images": images,
        "identity_ids": [s["identity_id"] for s in samples],
        "finger_ids": [s["finger_id"] for s in samples],
        "sensor_ids": [s["sensor_id"] for s in samples],
        "datasets": [s["dataset"] for s in samples],
        "image_paths": [s["image_path"] for s in samples],
    }