asdf / app.py
tkdehf2's picture
Update app.py
1adbe0b verified
raw
history blame
5.79 kB
import gradio as gr
from transformers import TrOCRProcessor, VisionEncoderDecoderModel
import torch
from PIL import Image
import numpy as np
import cv2
# OCR ๋ชจ๋ธ ๋ฐ ํ”„๋กœ์„ธ์„œ ์ดˆ๊ธฐํ™”
processor = TrOCRProcessor.from_pretrained('microsoft/trocr-base-handwritten')
model = VisionEncoderDecoderModel.from_pretrained('microsoft/trocr-base-handwritten')
# ์ •๋‹ต ๋ฐ ํ•ด์„ค ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค (20๋ฌธ์ œ)
answer_key = {
"1": {"answer": "๋ฏผ์ฃผ์ฃผ์˜", "explanation": "๋ฏผ์ฃผ์ฃผ์˜๋Š” ๊ตญ๋ฏผ์ด ์ฃผ์ธ์ด ๋˜์–ด ๋‚˜๋ผ์˜ ์ค‘์š”ํ•œ ์ผ์„ ๊ฒฐ์ •ํ•˜๋Š” ์ œ๋„์ž…๋‹ˆ๋‹ค."},
"2": {"answer": "์‚ผ๊ถŒ๋ถ„๋ฆฝ", "explanation": "์‚ผ๊ถŒ๋ถ„๋ฆฝ์€ ์ž…๋ฒ•๋ถ€, ํ–‰์ •๋ถ€, ์‚ฌ๋ฒ•๋ถ€๋กœ ๊ถŒ๋ ฅ์„ ๋‚˜๋ˆ„์–ด ์„œ๋กœ ๊ฒฌ์ œ์™€ ๊ท ํ˜•์„ ์ด๋ฃจ๊ฒŒ ํ•˜๋Š” ์ œ๋„์ž…๋‹ˆ๋‹ค."},
"3": {"answer": "์ง€๋ฐฉ์ž์น˜์ œ๋„", "explanation": "์ง€๋ฐฉ์ž์น˜์ œ๋„๋Š” ์ง€์—ญ์˜ ์ผ์„ ๊ทธ ์ง€์—ญ ์ฃผ๋ฏผ๋“ค์ด ์ง์ ‘ ๊ฒฐ์ •ํ•˜๊ณ  ์ฒ˜๋ฆฌํ•˜๋Š” ์ œ๋„์ž…๋‹ˆ๋‹ค."},
"4": {"answer": "ํ—Œ๋ฒ•", "explanation": "ํ—Œ๋ฒ•์€ ๊ตญ๊ฐ€์˜ ์ตœ๊ณ  ๋ฒ•์œผ๋กœ, ๊ตญ๋ฏผ์˜ ๊ธฐ๋ณธ๊ถŒ๊ณผ ์ •๋ถ€ ์กฐ์ง์— ๋Œ€ํ•œ ๊ธฐ๋ณธ ์›์น™์„ ๋‹ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค."},
"5": {"answer": "๊ตญํšŒ", "explanation": "๊ตญํšŒ๋Š” ๋ฒ•๋ฅ ์„ ๋งŒ๋“ค๊ณ  ์ •๋ถ€๋ฅผ ๊ฐ์‹œํ•˜๋Š” ์ž…๋ฒ•๋ถ€์˜ ์—ญํ• ์„ ๋‹ด๋‹นํ•ฉ๋‹ˆ๋‹ค."},
# 6~20๋ฒˆ๊นŒ์ง€ ๋ฌธ์ œ ์ถ”๊ฐ€ (์‹ค์ œ ์šด์˜ ์‹œ์—๋Š” ์—ฌ๊ธฐ์— ์ถ”๊ฐ€)
}
def segment_answers(image):
"""์‹œํ—˜์ง€์—์„œ ๋‹ต์•ˆ ์˜์—ญ์„ ๋ถ„ํ• ํ•˜๋Š” ํ•จ์ˆ˜"""
if isinstance(image, np.ndarray):
pil_image = Image.fromarray(image)
else:
return None
# ์ด๋ฏธ์ง€๋ฅผ ๊ทธ๋ ˆ์ด์Šค์ผ€์ผ๋กœ ๋ณ€ํ™˜
gray = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
# ์ด๋ฏธ์ง€ ์ด์ง„ํ™”
_, thresh = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY_INV)
# ์œค๊ณฝ์„  ์ฐพ๊ธฐ
contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
# ๋‹ต์•ˆ ์˜์—ญ ์ถ”์ถœ
answer_regions = []
for contour in contours:
x, y, w, h = cv2.boundingRect(contour)
if w > 50 and h > 20: # ์ตœ์†Œ ํฌ๊ธฐ ํ•„ํ„ฐ๋ง
region = image[y:y+h, x:x+w]
answer_regions.append({
'image': region,
'position': (y, x) # y์ขŒํ‘œ๋กœ ์ •๋ ฌํ•˜๊ธฐ ์œ„ํ•ด (y,x) ์ˆœ์„œ๋กœ ์ €์žฅ
})
# y์ขŒํ‘œ๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์ •๋ ฌ (์œ„์—์„œ ์•„๋ž˜๋กœ)
answer_regions.sort(key=lambda x: x['position'][0])
return [region['image'] for region in answer_regions]
def recognize_text(image):
"""์†๊ธ€์”จ ์ธ์‹ ํ•จ์ˆ˜"""
if isinstance(image, np.ndarray):
image = Image.fromarray(image)
pixel_values = processor(image, return_tensors="pt").pixel_values
with torch.no_grad():
generated_ids = model.generate(pixel_values)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
return generated_text
def grade_answer(question_number, student_answer):
"""๋‹ต์•ˆ ์ฑ„์  ํ•จ์ˆ˜"""
question_number = str(question_number)
if question_number not in answer_key:
return None
correct_answer = answer_key[question_number]["answer"]
explanation = answer_key[question_number]["explanation"]
# ๋‹ต์•ˆ ๋น„๊ต (๋„์–ด์“ฐ๊ธฐ, ๋Œ€์†Œ๋ฌธ์ž ๋ฌด์‹œ)
is_correct = student_answer.replace(" ", "").lower() == correct_answer.replace(" ", "").lower()
return {
"๋ฌธ์ œ๋ฒˆํ˜ธ": question_number,
"ํ•™์ƒ๋‹ต์•ˆ": student_answer,
"์ •๋‹ต์—ฌ๋ถ€": "O" if is_correct else "X",
"์ •๋‹ต": correct_answer,
"ํ•ด์„ค": explanation
}
def process_full_exam(image):
"""์ „์ฒด ์‹œํ—˜์ง€ ์ฒ˜๋ฆฌ ํ•จ์ˆ˜"""
if image is None or not isinstance(image, np.ndarray):
return "์‹œํ—˜์ง€ ์ด๋ฏธ์ง€๋ฅผ ์—…๋กœ๋“œํ•ด์ฃผ์„ธ์š”."
try:
# ๋‹ต์•ˆ ์˜์—ญ ๋ถ„ํ• 
answer_regions = segment_answers(image)
if not answer_regions:
return "๋‹ต์•ˆ ์˜์—ญ์„ ์ฐพ์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€๋ฅผ ํ™•์ธํ•ด์ฃผ์„ธ์š”."
# ์ฑ„์  ๊ฒฐ๊ณผ ์ €์žฅ
results = []
total_correct = 0
# ๊ฐ ๋‹ต์•ˆ ์˜์—ญ ์ฒ˜๋ฆฌ
for idx, region in enumerate(answer_regions, 1):
if idx > len(answer_key): # ์ •์˜๋œ ๋ฌธ์ œ ์ˆ˜๋ฅผ ์ดˆ๊ณผํ•˜๋ฉด ์ค‘๋‹จ
break
# ํ…์ŠคํŠธ ์ธ์‹
recognized_text = recognize_text(region)
# ์ฑ„์ 
result = grade_answer(idx, recognized_text)
if result:
results.append(result)
if result["์ •๋‹ต์—ฌ๋ถ€"] == "O":
total_correct += 1
# ๊ฒฐ๊ณผ ํฌ๋งทํŒ…
score = (total_correct / len(results)) * 100
output = f"์ด์ : {score:.1f}์  (20๋ฌธ์ œ ์ค‘ {total_correct}๊ฐœ ์ •๋‹ต)\n\n"
output += "=== ์ƒ์„ธ ์ฑ„์  ๊ฒฐ๊ณผ ===\n\n"
for result in results:
output += f"""
[{result['๋ฌธ์ œ๋ฒˆํ˜ธ']}๋ฒˆ] {'โœ“' if result['์ •๋‹ต์—ฌ๋ถ€']=='O' else 'โœ—'}
ํ•™์ƒ๋‹ต์•ˆ: {result['ํ•™์ƒ๋‹ต์•ˆ']}
์ •๋‹ต: {result['์ •๋‹ต']}
ํ•ด์„ค: {result['ํ•ด์„ค']}
"""
return output
except Exception as e:
return f"์ฒ˜๋ฆฌ ์ค‘ ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ–ˆ์Šต๋‹ˆ๋‹ค: {str(e)}"
# Gradio ์ธํ„ฐํŽ˜์ด์Šค ์ƒ์„ฑ
iface = gr.Interface(
fn=process_full_exam,
inputs=gr.Image(label="์‹œํ—˜์ง€ ์ด๋ฏธ์ง€๋ฅผ ์—…๋กœ๋“œํ•˜์„ธ์š”", type="numpy"),
outputs=gr.Textbox(label="์ฑ„์  ๊ฒฐ๊ณผ"),
title="์ดˆ๋“ฑํ•™๊ต ์‚ฌํšŒ ์‹œํ—˜์ง€ ์ฑ„์  ํ”„๋กœ๊ทธ๋žจ",
description="""
์ „์ฒด ์‹œํ—˜์ง€๋ฅผ ํ•œ ๋ฒˆ์— ์ฑ„์ ํ•˜๋Š” ํ”„๋กœ๊ทธ๋žจ์ž…๋‹ˆ๋‹ค.
์‹œํ—˜์ง€์˜ ๋‹ต์•ˆ์ด ์ž˜ ๋ณด์ด๋„๋ก ๊นจ๋—ํ•˜๊ฒŒ ์Šค์บ”ํ•˜๊ฑฐ๋‚˜ ์ดฌ์˜ํ•ด์ฃผ์„ธ์š”.
""",
examples=[], # ์˜ˆ์‹œ ์ด๋ฏธ์ง€๋ฅผ ์ถ”๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค
)
if __name__ == "__main__":
iface.launch()