File size: 5,410 Bytes
3abf0cf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
from __future__ import annotations

from pathlib import Path
from typing import Any

import numpy as np
import pandas as pd
from PIL import Image
from skimage.feature import hog, local_binary_pattern
from tqdm import tqdm

from .augmentations import augment_pil_for_classical
from .preprocessing import image_size_from_config, load_pil_image, pil_to_uint8_array, resize_image
from .utils import get_logger


LOGGER = get_logger(__name__)


def _prepare_gray_image(
    image: str | Path | Image.Image,
    config: dict[str, Any],
    augment_id: int = 0,
) -> np.ndarray:
    pil = load_pil_image(image, mode="L")
    if augment_id > 0:
        pil = augment_pil_for_classical(pil, augment_id, seed=int(config["seed"]))
    pil = resize_image(pil, image_size_from_config(config))
    return pil_to_uint8_array(pil)


def extract_hog_feature(
    image: str | Path | Image.Image,
    config: dict[str, Any],
    augment_id: int = 0,
) -> np.ndarray:
    gray = _prepare_gray_image(image, config, augment_id=augment_id)
    params = config["features"]["hog"]
    return hog(
        gray,
        orientations=int(params.get("orientations", 9)),
        pixels_per_cell=tuple(params.get("pixels_per_cell", [16, 16])),
        cells_per_block=tuple(params.get("cells_per_block", [2, 2])),
        block_norm=str(params.get("block_norm", "L2-Hys")),
        transform_sqrt=True,
        feature_vector=True,
    ).astype(np.float32)


def extract_lbp_feature(
    image: str | Path | Image.Image,
    config: dict[str, Any],
    augment_id: int = 0,
) -> np.ndarray:
    gray = _prepare_gray_image(image, config, augment_id=augment_id)
    params = config["features"]["lbp"]
    radius = int(params.get("radius", 2))
    n_points = int(params.get("n_points", 16))
    method = str(params.get("method", "uniform"))
    lbp = local_binary_pattern(gray, P=n_points, R=radius, method=method)
    if method == "uniform":
        n_bins = n_points + 2
    else:
        n_bins = min(2**n_points, 4096)
    hist, _ = np.histogram(lbp.ravel(), bins=n_bins, range=(0, n_bins), density=False)
    hist = hist.astype(np.float32)
    hist /= hist.sum() + 1e-8
    return hist


def extract_single_feature(
    image: str | Path | Image.Image,
    feature_type: str,
    config: dict[str, Any],
    augment_id: int = 0,
) -> np.ndarray:
    if feature_type == "hog":
        return extract_hog_feature(image, config, augment_id=augment_id)
    if feature_type == "lbp":
        return extract_lbp_feature(image, config, augment_id=augment_id)
    raise ValueError(f"Unsupported classical feature type: {feature_type}")


def expand_train_dataframe_for_balance(df: pd.DataFrame, config: dict[str, Any]) -> pd.DataFrame:
    if not config.get("balance", {}).get("enabled", True):
        return df.assign(augment_id=0, is_augmented=False)
    counts = df["label"].value_counts()
    if len(counts) < 2:
        return df.assign(augment_id=0, is_augmented=False)
    ratio = counts.max() / max(counts.min(), 1)
    threshold = float(config.get("data", {}).get("imbalance_threshold", 1.2))
    if ratio <= threshold:
        return df.assign(augment_id=0, is_augmented=False)

    rng = np.random.default_rng(int(config["seed"]))
    target = int(counts.max())
    rows = [df.assign(augment_id=0, is_augmented=False)]
    for label, count in counts.items():
        needed = target - int(count)
        if needed <= 0:
            continue
        candidates = df[df["label"] == label]
        sampled_idx = rng.choice(candidates.index.to_numpy(), size=needed, replace=True)
        augmented = candidates.loc[sampled_idx].copy().reset_index(drop=True)
        augmented["augment_id"] = np.arange(1, needed + 1)
        augmented["is_augmented"] = True
        rows.append(augmented)
        LOGGER.info("Classical balancing: added %d augmented '%s' training samples.", needed, label)
    expanded = pd.concat(rows, ignore_index=True)
    max_samples = int(config.get("balance", {}).get("max_augmented_train_samples", 3000) or 0)
    if max_samples > 0 and len(expanded) > max_samples:
        expanded = (
            expanded.groupby("label", group_keys=False)
            .sample(n=max_samples // expanded["label"].nunique(), random_state=int(config["seed"]), replace=False)
            .reset_index(drop=True)
        )
        LOGGER.info("Capped augmented classical training data at %d rows.", len(expanded))
    return expanded.sample(frac=1.0, random_state=int(config["seed"])).reset_index(drop=True)


def extract_feature_matrix(
    df: pd.DataFrame,
    feature_type: str,
    config: dict[str, Any],
    balance_train: bool = False,
) -> tuple[np.ndarray, np.ndarray, pd.DataFrame]:
    working = expand_train_dataframe_for_balance(df, config) if balance_train else df.assign(
        augment_id=0, is_augmented=False
    )
    features: list[np.ndarray] = []
    labels: list[int] = []
    iterator = tqdm(working.itertuples(index=False), total=len(working), desc=f"{feature_type.upper()} features")
    for row in iterator:
        features.append(
            extract_single_feature(
                getattr(row, "filepath"),
                feature_type,
                config,
                augment_id=int(getattr(row, "augment_id", 0)),
            )
        )
        labels.append(int(getattr(row, "label_id")))
    return np.vstack(features).astype(np.float32), np.asarray(labels, dtype=np.int64), working