7 Commits

13 changed files with 351 additions and 17 deletions

View File

@ -4,11 +4,10 @@ repos:
hooks: hooks:
- id: end-of-file-fixer - id: end-of-file-fixer
- id: trailing-whitespace - id: trailing-whitespace
- repo: https://github.com/psf/black
rev: 23.1.0 - repo: https://github.com/astral-sh/ruff-pre-commit
rev: v0.11.13
hooks: hooks:
- id: black - id: ruff
- repo: https://github.com/PyCQA/isort args: ["--fix"]
rev: 5.12.0 - id: ruff-format
hooks:
- id: isort

View File

@ -1,3 +1,2 @@
attrs==23.1.0 numpy~=2.3
numpy==1.26.1 opencv-python~=4.11
opencv-python==4.8.1.78

View File

@ -1,10 +1,9 @@
from dataclasses import dataclass
from datetime import datetime from datetime import datetime
from typing import Optional from typing import Optional
import attrs
@dataclass
@attrs.define
class B30OcrResultItem: class B30OcrResultItem:
rating_class: int rating_class: int
score: int score: int

View File

View File

@ -0,0 +1,3 @@
from .index import average, dct, difference
__all__ = ["average", "dct", "difference"]

View File

@ -0,0 +1,7 @@
import cv2
from arcaea_offline_ocr.types import Mat
def _resize_image(src: Mat, dsize: ...) -> Mat:
return cv2.resize(src, dsize, fx=0, fy=0, interpolation=cv2.INTER_AREA)

View File

@ -0,0 +1,35 @@
import cv2
import numpy as np
from arcaea_offline_ocr.types import Mat
from ._common import _resize_image
def average(img_gray: Mat, hash_size: int) -> Mat:
img_resized = _resize_image(img_gray, (hash_size, hash_size))
diff = img_resized > img_resized.mean()
return diff.flatten()
def difference(img_gray: Mat, hash_size: int) -> Mat:
img_size = (hash_size + 1, hash_size)
img_resized = _resize_image(img_gray, img_size)
previous = img_resized[:, :-1]
current = img_resized[:, 1:]
diff = previous > current
return diff.flatten()
def dct(img_gray: Mat, hash_size: int = 16, high_freq_factor: int = 4) -> Mat:
# TODO: consistency?
img_size_base = hash_size * high_freq_factor
img_size = (img_size_base, img_size_base)
img_resized = _resize_image(img_gray, img_size)
img_resized = img_resized.astype(np.float32)
dct_mat = cv2.dct(img_resized)
hash_mat = dct_mat[:hash_size, :hash_size]
return hash_mat > hash_mat.mean()

View File

@ -0,0 +1,18 @@
from .builder import ImageHashesDatabaseBuilder
from .index import ImageHashesDatabase, ImageHashesDatabasePropertyMissingError
from .models import (
ImageHashBuildTask,
ImageHashHashType,
ImageHashResult,
ImageHashCategory,
)
__all__ = [
"ImageHashesDatabase",
"ImageHashesDatabasePropertyMissingError",
"ImageHashHashType",
"ImageHashResult",
"ImageHashCategory",
"ImageHashesDatabaseBuilder",
"ImageHashBuildTask",
]

View File

@ -0,0 +1,85 @@
import logging
from datetime import datetime, timezone
from sqlite3 import Connection
from typing import List
from arcaea_offline_ocr.core import hashers
from .index import ImageHashesDatabase
from .models import ImageHash, ImageHashBuildTask, ImageHashHashType
logger = logging.getLogger(__name__)
class ImageHashesDatabaseBuilder:
@staticmethod
def __insert_property(conn: Connection, key: str, value: str):
return conn.execute(
"INSERT INTO properties (key, value) VALUES (?, ?)",
(key, value),
)
@classmethod
def build(
cls,
conn: Connection,
tasks: List[ImageHashBuildTask],
*,
hash_size: int = 16,
high_freq_factor: int = 4,
):
rows: List[ImageHash] = []
for task in tasks:
try:
img_gray = task.imread_function(task.image_path)
for hash_type, hash_mat in [
(
ImageHashHashType.AVERAGE,
hashers.average(img_gray, hash_size),
),
(
ImageHashHashType.DCT,
hashers.dct(img_gray, hash_size, high_freq_factor),
),
(
ImageHashHashType.DIFFERENCE,
hashers.difference(img_gray, hash_size),
),
]:
rows.append(
ImageHash(
hash_type=hash_type,
category=task.category,
label=task.label,
hash=ImageHashesDatabase.hash_mat_to_bytes(hash_mat),
)
)
except Exception:
logger.exception("Error processing task %r", task)
conn.execute("CREATE TABLE properties (`key` VARCHAR, `value` VARCHAR)")
conn.execute(
"CREATE TABLE hashes (`hash_type` INTEGER, `category` INTEGER, `label` VARCHAR, `hash` BLOB)"
)
now = datetime.now(tz=timezone.utc)
timestamp = int(now.timestamp() * 1000)
cls.__insert_property(conn, ImageHashesDatabase.KEY_HASH_SIZE, str(hash_size))
cls.__insert_property(
conn, ImageHashesDatabase.KEY_HIGH_FREQ_FACTOR, str(high_freq_factor)
)
cls.__insert_property(
conn, ImageHashesDatabase.KEY_BUILT_TIMESTAMP, str(timestamp)
)
conn.executemany(
"INSERT INTO hashes (hash_type, category, label, hash) VALUES (?, ?, ?, ?)",
[
(row.hash_type.value, row.category.value, row.label, row.hash)
for row in rows
],
)
conn.commit()

View File

@ -0,0 +1,144 @@
import sqlite3
from datetime import datetime, timezone
from typing import Any, Callable, List, Optional, TypeVar
from arcaea_offline_ocr.core import hashers
from arcaea_offline_ocr.types import Mat
from .models import ImageHashHashType, ImageHashResult, ImageHashCategory
T = TypeVar("T")
def _sql_hamming_distance(hash1: bytes, hash2: bytes):
assert len(hash1) == len(hash2), "hash size does not match!"
count = sum(1 for byte1, byte2 in zip(hash1, hash2) if byte1 != byte2)
return count
class ImageHashesDatabasePropertyMissingError(Exception):
pass
class ImageHashesDatabase:
KEY_HASH_SIZE = "hash_size"
KEY_HIGH_FREQ_FACTOR = "high_freq_factor"
KEY_BUILT_TIMESTAMP = "built_timestamp"
def __init__(self, conn: sqlite3.Connection):
self.conn = conn
self.conn.create_function("HAMMING_DISTANCE", 2, _sql_hamming_distance)
self._hash_size: int = -1
self._high_freq_factor: int = -1
self._built_time: Optional[datetime] = None
self._hashes_count = {
ImageHashCategory.JACKET: 0,
ImageHashCategory.PARTNER_ICON: 0,
}
self._hash_length: int = -1
self._initialize()
@property
def hash_size(self):
return self._hash_size
@property
def high_freq_factor(self):
return self._high_freq_factor
@property
def hash_length(self):
return self._hash_length
def _initialize(self):
def query_property(key, convert_func: Callable[[Any], T]) -> Optional[T]:
result = self.conn.execute(
"SELECT value FROM properties WHERE key = ?",
(key,),
).fetchone()
return convert_func(result[0]) if result is not None else None
def set_hashes_count(category: ImageHashCategory):
self._hashes_count[category] = self.conn.execute(
"SELECT COUNT(DISTINCT label) FROM hashes WHERE category = ?",
(category.value,),
).fetchone()[0]
hash_size = query_property(self.KEY_HASH_SIZE, lambda x: int(x))
if hash_size is None:
raise ImageHashesDatabasePropertyMissingError("hash_size")
self._hash_size = hash_size
high_freq_factor = query_property(self.KEY_HIGH_FREQ_FACTOR, lambda x: int(x))
if high_freq_factor is None:
raise ImageHashesDatabasePropertyMissingError("high_freq_factor")
self._high_freq_factor = high_freq_factor
self._built_time = query_property(
self.KEY_BUILT_TIMESTAMP,
lambda ts: datetime.fromtimestamp(int(ts) / 1000, tz=timezone.utc),
)
set_hashes_count(ImageHashCategory.JACKET)
set_hashes_count(ImageHashCategory.PARTNER_ICON)
self._hash_length = self._hash_size**2
def lookup_hash(
self, category: ImageHashCategory, hash_type: ImageHashHashType, hash: bytes
) -> List[ImageHashResult]:
cursor = self.conn.execute(
"SELECT"
" label,"
" HAMMING_DISTANCE(hash, ?) AS distance"
" FROM hashes"
" WHERE category = ? AND hash_type = ?"
" ORDER BY distance ASC LIMIT 10",
(hash, category.value, hash_type.value),
)
results = []
for label, distance in cursor.fetchall():
results.append(
ImageHashResult(
hash_type=hash_type,
category=category,
label=label,
confidence=(self.hash_length - distance) / self.hash_length,
)
)
return results
@staticmethod
def hash_mat_to_bytes(hash: Mat) -> bytes:
return bytes([255 if b else 0 for b in hash.flatten()])
def identify_image(self, category: ImageHashCategory, img) -> List[ImageHashResult]:
results = []
ahash = hashers.average(img, self.hash_size)
dhash = hashers.difference(img, self.hash_size)
phash = hashers.dct(img, self.hash_size, self.high_freq_factor)
results.extend(
self.lookup_hash(
category, ImageHashHashType.AVERAGE, self.hash_mat_to_bytes(ahash)
)
)
results.extend(
self.lookup_hash(
category, ImageHashHashType.DIFFERENCE, self.hash_mat_to_bytes(dhash)
)
)
results.extend(
self.lookup_hash(
category, ImageHashHashType.DCT, self.hash_mat_to_bytes(phash)
)
)
return results

View File

@ -0,0 +1,46 @@
import dataclasses
from enum import IntEnum
from typing import Callable
import cv2
from arcaea_offline_ocr.types import Mat
class ImageHashHashType(IntEnum):
AVERAGE = 0
DIFFERENCE = 1
DCT = 2
class ImageHashCategory(IntEnum):
JACKET = 0
PARTNER_ICON = 1
@dataclasses.dataclass
class ImageHash:
hash_type: ImageHashHashType
category: ImageHashCategory
label: str
hash: bytes
@dataclasses.dataclass
class ImageHashResult:
hash_type: ImageHashHashType
category: ImageHashCategory
label: str
confidence: float
def _default_imread_gray(image_path: str):
return cv2.cvtColor(cv2.imread(image_path, cv2.IMREAD_COLOR), cv2.COLOR_BGR2GRAY)
@dataclasses.dataclass
class ImageHashBuildTask:
image_path: str
category: ImageHashCategory
label: str
imread_function: Callable[[str], Mat] = _default_imread_gray

View File

@ -1,9 +1,8 @@
from dataclasses import dataclass
from typing import Optional from typing import Optional
import attrs
@dataclass
@attrs.define
class DeviceOcrResult: class DeviceOcrResult:
rating_class: int rating_class: int
pure: int pure: int

View File

@ -110,7 +110,7 @@ class DeviceOcr:
@staticmethod @staticmethod
def preprocess_char_icon(img_gray: Mat): def preprocess_char_icon(img_gray: Mat):
h, w = img_gray.shape[:2] h, w = img_gray.shape[:2]
img = cv2.copyMakeBorder(img_gray, w - h, 0, 0, 0, cv2.BORDER_REPLICATE) img = cv2.copyMakeBorder(img_gray, max(w - h, 0), 0, 0, 0, cv2.BORDER_REPLICATE)
h, w = img.shape[:2] h, w = img.shape[:2]
img = cv2.fillPoly( img = cv2.fillPoly(
img, img,