CHORE(merge): merge from develop
Some checks failed
Build Docker Image / build-and-push (push) Has been cancelled
CI / lint-and-test (push) Has been cancelled

- Initial FastAPI setup and OMR service
- K3S deployment configuration
- Memory limits for image processing
This commit is contained in:
2026-01-06 17:34:49 +09:00
parent 615fe6e574
commit b27daaa00c
19 changed files with 20 additions and 660 deletions

617
fastapi/main.py Normal file
View File

@@ -0,0 +1,617 @@
from fastapi import FastAPI, HTTPException, UploadFile, File, Form
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from typing import Dict
import json
import cv2
import numpy as np
app = FastAPI(
title="Joossam OMR Grading API",
description="OMR 채점을 위한 FastAPI 서비스",
version="1.0.0"
)
# CORS 설정 - Vercel에서 호스팅되는 프론트엔드 허용
app.add_middleware(
CORSMiddleware,
allow_origins=[
"https://joossameng.vercel.app",
"https://*.vercel.app",
"http://localhost:3000",
"http://localhost:3001",
],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
# --- 상수 및 설정 ---
TARGET_WIDTH = 2480
TARGET_HEIGHT = 3508
# --- 함수 정의 ---
def load_image_from_bytes(image_bytes: bytes):
"""바이트 데이터에서 이미지 로드"""
nparr = np.frombuffer(image_bytes, np.uint8)
img = cv2.imdecode(nparr, cv2.IMREAD_COLOR)
if img is None:
raise ValueError("이미지를 읽을 수 없습니다")
return img, img.shape[:2]
def resize_image_to_target(img, target_width=TARGET_WIDTH, target_height=TARGET_HEIGHT):
"""이미지를 타겟 크기로 리사이징 (비율 유지)"""
h, w = img.shape[:2]
if w == target_width and h == target_height:
return img, 1.0, 1.0
scale_x = target_width / w
scale_y = target_height / h
new_width = int(w * scale_x)
new_height = int(h * scale_y)
resized_img = cv2.resize(img, (new_width, new_height), interpolation=cv2.INTER_CUBIC)
return resized_img, scale_x, scale_y
def gamma_correction(img, gamma=0.7):
"""감마 보정으로 밝기 곡선 조정"""
inv_gamma = 1.0 / gamma
table = np.array([((i / 255.0) ** inv_gamma) * 255 for i in np.arange(0, 256)]).astype("uint8")
return cv2.LUT(img, table)
def unsharp_mask(img, kernel_size=(5, 5), sigma=1.0, amount=1.0):
"""언샤프 마스킹으로 경계 선명화"""
if len(img.shape) == 3:
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
blurred = cv2.GaussianBlur(img, kernel_size, sigma)
sharpened = cv2.addWeighted(img, 1.0 + amount, blurred, -amount, 0)
return sharpened
def deskew_image_with_barcodes(img):
"""바코드를 기준으로 이미지 기울기 보정"""
try:
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
_, thresh = cv2.threshold(gray, 180, 255, cv2.THRESH_BINARY_INV)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
thresh = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, kernel)
thresh = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel)
contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
img_height, img_width = img.shape[:2]
top_area_threshold = img_height * 0.15
top_rectangles = []
for contour in contours:
x, y, w, h = cv2.boundingRect(contour)
if (y < top_area_threshold and
w > 10 and h > 10 and
w < 300 and h < 300):
center_x = x + w // 2
center_y = y + h // 2
top_rectangles.append({'center': (center_x, center_y)})
if len(top_rectangles) < 15:
return img
top_rectangles.sort(key=lambda x: x['center'][1])
top_rectangles = top_rectangles[:23]
top_rectangles.sort(key=lambda x: x['center'][0])
if len(top_rectangles) >= 10:
left_points = top_rectangles[:5]
right_points = top_rectangles[-5:]
left_avg_x = np.mean([p['center'][0] for p in left_points])
left_avg_y = np.mean([p['center'][1] for p in left_points])
right_avg_x = np.mean([p['center'][0] for p in right_points])
right_avg_y = np.mean([p['center'][1] for p in right_points])
delta_y = right_avg_y - left_avg_y
delta_x = right_avg_x - left_avg_x
if delta_x == 0:
return img
angle_rad = np.arctan2(delta_y, delta_x)
angle_deg = np.degrees(angle_rad)
if abs(angle_deg) < 0.3:
return img
if abs(angle_deg) > 10:
return img
center = (img_width // 2, img_height // 2)
rotation_matrix = cv2.getRotationMatrix2D(center, angle_deg, 1.0)
deskewed = cv2.warpAffine(
img, rotation_matrix, (img_width, img_height),
flags=cv2.INTER_CUBIC,
borderMode=cv2.BORDER_REPLICATE
)
return deskewed
return img
except Exception:
return img
def preprocess_omr_image(img):
"""OMR 이미지 전처리 강화"""
denoised = cv2.GaussianBlur(img, (3, 3), 0)
gamma_corrected = gamma_correction(denoised, gamma=0.7)
if len(gamma_corrected.shape) == 3:
gray = cv2.cvtColor(gamma_corrected, cv2.COLOR_BGR2GRAY)
else:
gray = gamma_corrected
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
enhanced = clahe.apply(gray)
sharpened = unsharp_mask(enhanced, amount=0.8)
result = cv2.cvtColor(sharpened, cv2.COLOR_GRAY2BGR)
return result
def calculate_marking_density(img, x, y, width=30, height=60):
"""특정 좌표 주변 영역의 마킹 밀도 계산"""
h, w = img.shape[:2]
x1 = max(0, x - width//2)
y1 = max(0, y - height//2)
x2 = min(w, x + width//2)
y2 = min(h, y + height//2)
region = img[y1:y2, x1:x2]
if region.size == 0:
return 0.0
if len(region.shape) == 3:
region = cv2.cvtColor(region, cv2.COLOR_BGR2GRAY)
avg_darkness = 255 - np.mean(region)
dark_pixels = np.sum(region < 180)
total_pixels = region.size
dark_ratio = dark_pixels / total_pixels
medium_dark_pixels = np.sum(region < 160)
medium_dark_ratio = medium_dark_pixels / total_pixels
very_dark_pixels = np.sum(region < 120)
very_dark_ratio = very_dark_pixels / total_pixels
density_score = (avg_darkness / 255.0) * 0.2 + dark_ratio * 0.2 + medium_dark_ratio * 0.4 + very_dark_ratio * 0.2
return density_score
upperValueSquare = 180
def find_top_black_rectangles(img):
"""상단 검은색 사각형들을 찾아서 좌표 반환"""
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
_, thresh = cv2.threshold(gray, upperValueSquare, 255, cv2.THRESH_BINARY_INV)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
thresh = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, kernel)
thresh = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel)
contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
top_rectangles = []
img_height = img.shape[0]
top_area_threshold = img_height * 0.15
for contour in contours:
x, y, w, h = cv2.boundingRect(contour)
if (y < top_area_threshold and
w > 10 and h > 10 and
w < 300 and h < 300):
center_x = x + w // 2
center_y = y + h // 2
top_rectangles.append({
'center': (center_x, center_y),
'bbox': (x, y, w, h),
'area': w * h
})
top_rectangles.sort(key=lambda x: x['center'][1])
selected_rectangles = top_rectangles[:23]
selected_rectangles.sort(key=lambda x: x['center'][0])
return selected_rectangles
def find_side_black_rectangles(img):
"""좌우측 검은색 사각형들을 찾아서 좌표 반환 (Y축 계산용)"""
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
_, thresh = cv2.threshold(gray, upperValueSquare, 255, cv2.THRESH_BINARY_INV)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
thresh = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, kernel)
thresh = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel)
contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
img_height, img_width = img.shape[:2]
left_area_threshold = img_width * 0.15
right_area_start = img_width * 0.85
left_rectangles = []
right_rectangles = []
for contour in contours:
x, y, w, h = cv2.boundingRect(contour)
center_x = x + w // 2
center_y = y + h // 2
if (w > 20 and h > 20 and
w < 70 and h < 70):
if center_x < left_area_threshold:
left_rectangles.append({
'center': (center_x, center_y),
'bbox': (x, y, w, h),
'area': w * h
})
elif center_x > right_area_start:
right_rectangles.append({
'center': (center_x, center_y),
'bbox': (x, y, w, h),
'area': w * h
})
left_rectangles.sort(key=lambda x: x['center'][0])
left_selected = left_rectangles[:10]
left_selected.sort(key=lambda x: x['center'][1])
right_rectangles.sort(key=lambda x: x['center'][0])
right_selected = right_rectangles[-20:] if len(right_rectangles) >= 20 else right_rectangles
right_selected.sort(key=lambda x: x['center'][1])
return left_selected, right_selected
def define_phone_positions(img_width, img_height, top_rectangles, left_rectangles):
"""전화번호 전용 위치 계산 (1-8번 사각형, 1-10번 숫자 자리)"""
phone_positions = {}
for rect_index in range(8):
if rect_index < len(top_rectangles):
rect_center_x = top_rectangles[rect_index]['center'][0]
digit_position = rect_index + 1
if len(left_rectangles) >= 10:
phone_positions[digit_position] = {}
for digit in range(10):
if digit < len(left_rectangles):
y = left_rectangles[digit]['center'][1]
phone_positions[digit_position][str(digit)] = (rect_center_x, y)
return phone_positions
def define_answer_positions(img_width, img_height, top_rectangles, left_rectangles, right_rectangles):
"""답안 전용 위치 계산 (9-23번 사각형, 1-45번 문제)"""
positions = {}
for rect_index in range(8, 13):
if rect_index < len(top_rectangles):
rect_center_x = top_rectangles[rect_index]['center'][0]
choice_num = rect_index - 7
if len(right_rectangles) >= 20:
for q in range(1, 21):
if q not in positions:
positions[q] = {}
question_index = q - 1
if question_index < len(right_rectangles):
y = right_rectangles[question_index]['center'][1]
positions[q][str(choice_num)] = (rect_center_x, y)
for rect_index in range(13, 18):
if rect_index < len(top_rectangles):
rect_center_x = top_rectangles[rect_index]['center'][0]
choice_num = rect_index - 12
if len(right_rectangles) >= 20:
for q in range(21, 41):
if q not in positions:
positions[q] = {}
question_index = q - 21
if question_index < len(right_rectangles):
y = right_rectangles[question_index]['center'][1]
positions[q][str(choice_num)] = (rect_center_x, y)
for rect_index in range(18, 23):
if rect_index < len(top_rectangles):
rect_center_x = top_rectangles[rect_index]['center'][0]
choice_num = rect_index - 17
if len(right_rectangles) >= 5:
for q in range(41, 46):
if q not in positions:
positions[q] = {}
question_index = q - 41
if question_index < len(right_rectangles):
y = right_rectangles[question_index]['center'][1]
positions[q][str(choice_num)] = (rect_center_x, y)
return positions
def estimate_phone_number_with_density(img, phone_positions, min_density=0.17):
"""전화번호 추정"""
phone_selected = {}
for digit_pos, digit_choices in phone_positions.items():
if not digit_choices:
phone_selected[digit_pos] = "0"
continue
digit_densities = {}
for digit, coord in digit_choices.items():
x, y = coord
density = calculate_marking_density(img, x, y)
digit_densities[digit] = density
highest_digit, highest_density = max(digit_densities.items(), key=lambda x: x[1])
if highest_density >= min_density:
phone_selected[digit_pos] = highest_digit
else:
phone_selected[digit_pos] = "0"
return phone_selected
def estimate_selected_answers_with_density(img, answer_positions, min_density=0.2):
"""답안 추정"""
selected = {}
for q_num, choices in answer_positions.items():
if not choices:
selected[str(q_num)] = "무효"
continue
choice_densities = {}
for choice, coord in choices.items():
x, y = coord
density = calculate_marking_density(img, x, y)
choice_densities[choice] = density
highest_choice, highest_density = max(choice_densities.items(), key=lambda x: x[1])
if highest_density >= min_density:
selected[str(q_num)] = highest_choice
else:
selected[str(q_num)] = "무효"
return selected
def extract_phone_number(phone_selected):
"""전화번호 8자리 추출"""
phone_digits = []
for i in range(1, 9):
if i in phone_selected:
digit = phone_selected[i]
if digit and digit != "무효":
try:
digit_int = int(digit)
if 0 <= digit_int <= 9:
phone_digits.append(str(digit_int))
else:
phone_digits.append("0")
except ValueError:
phone_digits.append("0")
else:
phone_digits.append("0")
else:
phone_digits.append("0")
phone_number = "".join(phone_digits)
return phone_number
def calculate_total_score(selected_answers, correct_answers, question_scores):
"""총점 계산"""
total = 0
for q_num, correct_answer in correct_answers.items():
if q_num in selected_answers:
student_answer = selected_answers[q_num]
if student_answer == correct_answer:
score = question_scores.get(q_num, 0)
total += score
return total
def calculate_grade(total_score):
"""등급 계산"""
if total_score >= 90:
return 1
elif total_score >= 80:
return 2
elif total_score >= 70:
return 3
elif total_score >= 60:
return 4
elif total_score >= 50:
return 5
elif total_score >= 40:
return 6
elif total_score >= 30:
return 7
elif total_score >= 20:
return 8
else:
return 9
def create_results_array(selected_answers, correct_answers, question_scores, question_types):
"""결과 배열 생성"""
results = []
for q_num in sorted(correct_answers.keys(), key=lambda x: int(x)):
try:
q_num_int = int(q_num)
student_answer = selected_answers.get(q_num, "무효")
correct_answer = correct_answers[q_num]
score = question_scores.get(q_num, 0)
question_type = question_types.get(q_num, "기타")
earned_score = score if student_answer == correct_answer else 0
results.append({
"questionNumber": q_num_int,
"studentAnswer": str(student_answer),
"correctAnswer": correct_answer,
"score": score,
"earnedScore": earned_score,
"questionType": question_type
})
except (ValueError, TypeError):
continue
return results
def grade_omr_from_bytes(image_bytes: bytes, correct_answers: Dict, question_scores: Dict, question_types: Dict):
"""OMR 채점 메인 함수 (바이트 입력)"""
try:
img, (h, w) = load_image_from_bytes(image_bytes)
deskewed_img = deskew_image_with_barcodes(img)
expected_ratio = TARGET_WIDTH / TARGET_HEIGHT
actual_ratio = w / h
resized_img, scale_x, scale_y = resize_image_to_target(deskewed_img)
resized_h, resized_w = resized_img.shape[:2]
preprocessed_img = preprocess_omr_image(resized_img)
top_rectangles = find_top_black_rectangles(resized_img)
left_rectangles, right_rectangles = find_side_black_rectangles(resized_img)
phone_positions = define_phone_positions(resized_w, resized_h, top_rectangles, left_rectangles)
answer_positions = define_answer_positions(resized_w, resized_h, top_rectangles, left_rectangles, right_rectangles)
phone_selected = estimate_phone_number_with_density(preprocessed_img, phone_positions)
phone_number = extract_phone_number(phone_selected)
selected_answers = estimate_selected_answers_with_density(preprocessed_img, answer_positions)
correct_count = 0
for q_num, correct_answer in correct_answers.items():
if q_num in selected_answers:
student_answer = selected_answers[q_num]
if str(student_answer) == str(correct_answer):
correct_count += 1
total_score = calculate_total_score(selected_answers, correct_answers, question_scores)
grade = calculate_grade(total_score)
results = create_results_array(selected_answers, correct_answers, question_scores, question_types)
final_result = {
"totalScore": total_score,
"grade": grade,
"phoneNumber": phone_number,
"results": results,
"imageInfo": {
"originalSize": f"{w}x{h}",
"resizedSize": f"{resized_w}x{resized_h}",
"scaleFactors": {"x": scale_x, "y": scale_y},
"aspectRatio": {"expected": expected_ratio, "actual": actual_ratio}
}
}
return final_result
except Exception as e:
raise HTTPException(status_code=500, detail=f"OMR 채점 실패: {str(e)}")
# API 엔드포인트들
@app.get("/")
async def root():
"""헬스체크 엔드포인트"""
return {"status": "healthy", "message": "Joossam OMR Grading API is running"}
@app.get("/health")
async def health():
"""헬스체크 엔드포인트"""
return {"status": "healthy"}
class GradingRequest(BaseModel):
correct_answers: Dict[str, str]
question_scores: Dict[str, int]
question_types: Dict[str, str]
@app.post("/api/omr/grade")
async def grade_omr(
image: UploadFile = File(...),
correct_answers: str = Form(...),
question_scores: str = Form(...),
question_types: str = Form(...)
):
"""
OMR 채점 API
- image: OMR 이미지 파일
- correct_answers: 정답 JSON (예: {"1": "3", "2": "1", ...})
- question_scores: 문제별 점수 JSON (예: {"1": 2, "2": 2, ...})
- question_types: 문제 유형 JSON (예: {"1": "어휘", "2": "문법", ...})
"""
try:
# JSON 파싱
correct_answers_dict = json.loads(correct_answers)
question_scores_dict = json.loads(question_scores)
question_types_dict = json.loads(question_types)
# 이미지 읽기
image_bytes = await image.read()
# OMR 채점 실행
result = grade_omr_from_bytes(
image_bytes,
correct_answers_dict,
question_scores_dict,
question_types_dict
)
return result
except json.JSONDecodeError as e:
raise HTTPException(status_code=400, detail=f"JSON 파싱 오류: {str(e)}")
except Exception as e:
raise HTTPException(status_code=500, detail=f"채점 오류: {str(e)}")
if __name__ == "__main__":
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=8000)

8
fastapi/requirements.txt Normal file
View File

@@ -0,0 +1,8 @@
fastapi==0.109.0
uvicorn[standard]==0.27.0
python-multipart==0.0.6
opencv-python-headless==4.8.1.78
numpy==1.24.3
Pillow==10.3.0
pydantic==2.5.3