import cv2
import numpy as np
import pandas as pd
from typing import Dict, List, Tuple, Optional
class NewGaokaoAnswerSheetRecognizer:
def __init__(self):
# 新高考科目配置示例 (3+1+2模式)
self.subject_config = {
'common': ['chinese', 'math', 'english'], # 语数英(所有考生必考)
'preferred': ['physics', 'history'], # 首选科目(二选一)
'elective': { # 再选科目(四选二)
'physics_group': ['chemistry', 'biology', 'geography', 'politics'],
'history_group': ['chemistry', 'biology', 'geography', 'politics']
}
}
# 答题卡模板参数(需根据实际答题卡调整)
self.template_params = {
'定位点': {
'数量': 4,
'尺寸_mm': 8, # 8mm x 8mm
'位置': ['左上', '右上', '左下', '右下']
},
'考号区域': {
'行数': 9, # 9位考号
'列数': 10, # 0-9
'起始坐标': (100, 50) # 示例坐标
},
'科目识别码区域': {
'位置': (50, 150), # 用于识别物理组/历史组
'尺寸': (100, 30)
}
}
self.recognized_data = {
'student_id': '',
'subject_group': '',
'answers': {},
'raw_fill_values': {}
}
def preprocess_image(self, image_path: str, is_backside: bool = False) -> Tuple[np.ndarray, np.ndarray]:
"""
图像预处理:支持正反面的独立处理
Args:
image_path: 图像路径
is_backside: 是否是背面(有些答题卡背面有不同布局)
Returns:
original: 原始图像
binary: 预处理后的二值图像
"""
# 读取图像
img = cv2.imread(image_path)
if img is None:
raise FileNotFoundError(f"无法读取图像: {image_path}")
# 1. 灰度化 - 使用人眼感知加权公式 [citation:5]
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# 2. 自适应阈值二值化 - 解决光照不均问题
# blockSize=15, C=2 是经过大量测试的黄金参数 [citation:5]
binary = cv2.adaptiveThreshold(
gray, 255,
cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY_INV, # 填涂区域变白,便于分析
15, 2
)
# 3. 形态学操作 - 去除噪声,连接断裂线条
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))
binary = cv2.morphologyEx(binary, cv2.MORPH_CLOSE, kernel) # 闭运算补线
binary = cv2.medianBlur(binary, 3) # 中值滤波去椒盐噪声 [citation:5]
return img, binary
def locate_answer_sheet(self, binary_img: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
"""
定位答题卡并执行透视校正
Returns:
warped: 校正后的图像
transform_matrix: 透视变换矩阵
"""
# 1. 查找轮廓
contours, _ = cv2.findContours(
binary_img,
cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE
)
# 2. 筛选最大的四边形(应该是答题卡轮廓)
max_area = 0
sheet_contour = None
for contour in contours:
area = cv2.contourArea(contour)
if area < 10000: # 过滤太小的轮廓
continue
# 多边形近似
epsilon = 0.02 * cv2.arcLength(contour, True)
approx = cv2.approxPolyDP(contour, epsilon, True)
# 如果是四边形且面积最大
if len(approx) == 4 and area > max_area:
max_area = area
sheet_contour = approx
if sheet_contour is None:
raise ValueError("未找到答题卡轮廓")
# 3. 获取四个角点并排序 (左上、右上、右下、左下)
pts = sheet_contour.reshape(4, 2)
rect = np.zeros((4, 2), dtype=np.float32)
# 计算中心点
center = np.mean(pts, axis=0)
# 根据角度排序
angles = []
for pt in pts:
angle = np.arctan2(pt[1] - center[1], pt[0] - center[0])
angles.append(angle)
sorted_idx = np.argsort(angles)
pts = pts[sorted_idx]
# 调整顺序为:左上、右上、右下、左下
rect[0] = pts[0] # 左上
rect[1] = pts[1] # 右上
rect[2] = pts[2] # 右下
rect[3] = pts[3] # 左下
# 4. 计算目标尺寸
width = max(
np.linalg.norm(rect[1] - rect[0]),
np.linalg.norm(rect[2] - rect[3])
)
height = max(
np.linalg.norm(rect[3] - rect[0]),
np.linalg.norm(rect[2] - rect[1])
)
# 5. 执行透视变换 [citation:5]
dst_pts = np.array([
[0, 0],
[width - 1, 0],
[width - 1, height - 1],
[0, height - 1]
], dtype=np.float32)
transform_matrix = cv2.getPerspectiveTransform(rect, dst_pts)
warped = cv2.warpPerspective(
binary_img, transform_matrix,
(int(width), int(height))
)
return warped, transform_matrix
def locate_registration_marks(self, binary_img: np.ndarray) -> Dict[str, Tuple[int, int]]:
"""
定位答题卡四角的定位标记(黑块)
这些标记用于精确校正和区域划分 [citation:5]
"""
# 使用模板匹配查找定位点
marks = {}
# 创建定位点模板(8x8mm的黑色方块)
template_size = 30 # 像素,需根据实际DPI调整
template = np.zeros((template_size, template_size), dtype=np.uint8)
cv2.rectangle(template, (2, 2), (template_size-2, template_size-2), 255, -1)
# 在图像四角区域搜索
h, w = binary_img.shape
search_regions = {
'top_left': (0, 0, w//4, h//4),
'top_right': (w*3//4, 0, w//4, h//4),
'bottom_left': (0, h*3//4, w//4, h//4),
'bottom_right': (w*3//4, h*3//4, w//4, h//4)
}
for name, (x, y, rw, rh) in search_regions.items():
roi = binary_img[y:y+rh, x:x+rw]
if roi.size == 0:
continue
# 模板匹配
result = cv2.matchTemplate(roi, template, cv2.TM_CCOEFF_NORMED)
_, max_val, _, max_loc = cv2.minMaxLoc(result)
if max_val > 0.6: # 匹配度阈值
marks[name] = (x + max_loc[0] + template_size//2,
y + max_loc[1] + template_size//2)
return marks
def extract_student_id(self, binary_img: np.ndarray) -> str:
"""
识别学生考号(9位数字填涂)
使用行列投影法识别填涂 [citation:6]
"""
h, w = binary_img.shape
# 考号区域坐标(需根据实际答题卡标定)
id_x_start, id_y_start = 200, 100
id_width, id_height = 400, 300
id_region = binary_img[id_y_start:id_y_start+id_height,
id_x_start:id_x_start+id_width]
# 划分9行(数字位)和10列(0-9选项)
digit_height = id_height // 9
digit_width = id_width // 10
student_id = ''
fill_threshold = 0.64 # 填涂面积阈值(64%面积被涂黑算填涂)[citation:6]
for row in range(9): # 9位考号
max_fill_ratio = 0
selected_digit = None
for col in range(10): # 每个数字位的10个选项
# 提取单个选项区域
x = col * digit_width
y = row * digit_height
option_region = id_region[y:y+digit_height, x:x+digit_width]
# 计算填涂比例
fill_ratio = np.sum(option_region == 255) / option_region.size
if fill_ratio > max_fill_ratio:
max_fill_ratio = fill_ratio
selected_digit = col
# 判断是否有效填涂
if max_fill_ratio >= fill_threshold:
student_id += str(selected_digit)
else:
student_id += '?' # 未填涂或模糊
return student_id
def detect_subject_group(self, binary_img: np.ndarray) -> str:
"""
识别考生选择的科目组(物理组/历史组)
通过科目识别码区域判断 [citation:8]
"""
# 科目识别码区域(通常是一个特殊标记或填涂区)
x, y = self.template_params['科目识别码区域']['位置']
w, h = self.template_params['科目识别码区域']['尺寸']
code_region = binary_img[y:y+h, x:x+w]
# 简单的二值化判断
mean_value = np.mean(code_region)
# 根据识别码判断科目组(实际应用中需要更复杂的逻辑)
if mean_value > 127:
return 'physics_group' # 物理组
else:
return 'history_group' # 历史组
def extract_answers(self, binary_img: np.ndarray, subject_group: str) -> Dict[str, List[str]]:
"""
提取所有题目的填涂答案
支持多科目合卡,根据科目组选择需要识别的区域 [citation:7]
"""
answers = {}
# 根据科目组确定要识别的科目
subjects = ['chinese', 'math', 'english'] # 语数英必考
if subject_group == 'physics_group':
subjects.append('physics') # 首选科目
# 再选科目(4选2)- 需要特殊处理,通常会有4个填涂区选2个
subjects.extend(['elective1', 'elective2'])
else: # history_group
subjects.append('history')
subjects.extend(['elective1', 'elective2'])
# 各科目答题区域坐标(需根据实际答题卡标定)
# 这里使用示例坐标,实际应用中需要通过模板文件配置
subject_regions = {
'chinese': {'x': 50, 'y': 200, 'w': 500, 'h': 400, 'questions': 15, 'options': 4},
'math': {'x': 50, 'y': 650, 'w': 500, 'h': 400, 'questions': 12, 'options': 4},
'english': {'x': 600, 'y': 200, 'w': 500, 'h': 400, 'questions': 15, 'options': 4},
'physics': {'x': 600, 'y': 650, 'w': 500, 'h': 400, 'questions': 10, 'options': 4},
'history': {'x': 600, 'y': 650, 'w': 500, 'h': 400, 'questions': 10, 'options': 4},
'elective1': {'x': 50, 'y': 1100, 'w': 500, 'h': 200, 'questions': 5, 'options': 4},
'elective2': {'x': 600, 'y': 1100, 'w': 500, 'h': 200, 'questions': 5, 'options': 4}
}
# 对每个科目提取答案
for subject in subjects:
if subject not in subject_regions:
continue
region = subject_regions[subject]
roi = binary_img[region['y']:region['y']+region['h'],
region['x']:region['x']+region['w']]
# 计算每个题目的每个选项区域
q_height = region['h'] // region['questions']
opt_width = region['w'] // region['options']
subject_answers = []
for q in range(region['questions']):
filled_options = []
for opt in range(region['options']):
# 提取选项区域
x = opt * opt_width
y = q * q_height
opt_region = roi[y:y+q_height, x:x+opt_width]
# 计算填涂比例
fill_ratio = np.sum(opt_region == 255) / opt_region.size
# 填涂判断阈值,可根据实际情况调整 [citation:6]
if fill_ratio > 0.4:
filled_options.append(chr(65 + opt)) # A, B, C, D...
subject_answers.append(''.join(filled_options) if filled_options else '')
answers[subject] = subject_answers
return answers
def calculate_score(self, answers: Dict[str, List[str]],
answer_key: Dict[str, List[str]]) -> Dict[str, float]:
"""
计算得分(支持赋分制)
Args:
answers: 识别出的学生答案
answer_key: 标准答案
Returns:
各科目得分和原始总分
"""
scores = {}
total_raw_score = 0
for subject, student_ans in answers.items():
if subject not in answer_key:
continue
key = answer_key[subject]
subject_score = 0
# 逐题比对
for i, (s_ans, k_ans) in enumerate(zip(student_ans, key)):
if s_ans and s_ans == k_ans: # 单选情况
# 假设每题3分
subject_score += 3
elif s_ans and set(s_ans) == set(k_ans): # 多选全对
subject_score += 3
# 可根据实际规则增加部分得分逻辑
scores[subject] = subject_score
total_raw_score += subject_score
scores['raw_total'] = total_raw_score
# 如果是再选科目,需要进行赋分转换(示例逻辑)
if 'elective1' in scores:
# 根据原始分在全省(或本校)的排名进行赋分
# 这里简化处理:原始分 -> 等级赋分(100分制)
scores['elective1_converted'] = self.raw_to_rank_score(scores['elective1'], 3, 5)
scores['elective2_converted'] = self.raw_to_rank_score(scores['elective2'], 3, 5)
return scores
def raw_to_rank_score(self, raw_score: float, question_count: int,
score_per_question: int) -> int:
"""
原始分转等级赋分(模拟)
新高考再选科目采用等级赋分制 [citation:3][citation:8]
"""
# 这是一个简化的示例,实际需要根据全省排名
max_possible = question_count * score_per_question
percentage = raw_score / max_possible if max_possible > 0 else 0
# 假设的赋分转换表
if percentage >= 0.9:
return 100
elif percentage >= 0.8:
return 90
elif percentage >= 0.7:
return 80
elif percentage >= 0.6:
return 70
elif percentage >= 0.5:
return 60
else:
return 50
def process_answer_sheet(self, front_image: str, back_image: Optional[str] = None,
answer_key: Optional[Dict[str, List[str]]] = None) -> Dict:
"""
主处理流程:支持正反面答题卡 [citation:3][citation:8]
"""
print("=" * 60)
print("新高考答题卡识别系统启动")
print("=" * 60)
result = {
'student_id': '',
'subject_group': '',
'answers': {},
'scores': {},
'images': {},
'status': 'success'
}
try:
# 1. 处理正面
print("\n1. 处理答题卡正面...")
img_front, binary_front = self.preprocess_image(front_image, is_backside=False)
# 2. 定位并校正
print("2. 图像定位与校正...")
warped_front, _ = self.locate_answer_sheet(binary_front)
# 3. 识别考号
print("3. 识别考生考号...")
student_id = self.extract_student_id(warped_front)
result['student_id'] = student_id
print(f" 考号: {student_id}")
# 4. 识别科目组
print("4. 识别科目组...")
subject_group = self.detect_subject_group(warped_front)
result['subject_group'] = subject_group
print(f" 科目组: {subject_group}")
# 5. 提取答案
print("5. 提取填涂答案...")
answers = self.extract_answers(warped_front, subject_group)
result['answers'] = answers
# 如果有背面,继续处理
if back_image:
print("\n6. 处理答题卡背面...")
img_back, binary_back = self.preprocess_image(back_image, is_backside=True)
warped_back, _ = self.locate_answer_sheet(binary_back)
back_answers = self.extract_answers(warped_back, subject_group)
# 合并正反面答案
for subject, ans in back_answers.items():
if subject not in answers:
answers[subject] = ans
# 6. 计算得分
if answer_key:
print("\n7. 计算得分...")
scores = self.calculate_score(answers, answer_key)
result['scores'] = scores
print(f" 原始总分: {scores.get('raw_total', 0)}")
if 'elective1_converted' in scores:
print(f" 再选科目1赋分: {scores['elective1_converted']}")
print(f" 再选科目2赋分: {scores['elective2_converted']}")
# 保存中间图像用于调试
result['images']['front_corrected'] = warped_front
print("\n" + "=" * 60)
print("识别完成!")
print("=" * 60)
except Exception as e:
result['status'] = 'error'
result['error_message'] = str(e)
print(f"\n处理出错: {e}")
return result
def export_to_excel(self, result: Dict, output_path: str):
"""导出识别结果到Excel"""
# 构建DataFrame
data = {
'考号': [result['student_id']],
'科目组': [result['subject_group']],
'识别状态': [result['status']]
}
# 添加各科目答案
for subject, answers in result.get('answers', {}).items():
data[f'{subject}_答案'] = [' '.join(answers)]
# 添加得分
for subject, score in result.get('scores', {}).items():
data[f'{subject}_得分'] = [score]
df = pd.DataFrame(data)
df.to_excel(output_path, index=False)
print(f"\n结果已导出到: {output_path}")
# ==================== 使用示例 ====================
if __name__ == "__main__":
# 初始化识别器
recognizer = NewGaokaoAnswerSheetRecognizer()
# 标准答案示例(新高考物理组)
answer_key = {
'chinese': ['A', 'B', 'C', 'D', 'A', 'B', 'C', 'D', 'A', 'B', 'C', 'D', 'A', 'B', 'C'],
'math': ['A', 'B', 'C', 'D', 'A', 'B', 'C', 'D', 'A', 'B', 'C', 'D'],
'english': ['A', 'B', 'C', 'D', 'A', 'B', 'C', 'D', 'A', 'B', 'C', 'D', 'A', 'B', 'C'],
'physics': ['A', 'B', 'C', 'D', 'A', 'B', 'C', 'D', 'A', 'B'],
'elective1': ['A', 'B', 'C', 'D', 'A'],
'elective2': ['B', 'C', 'D', 'A', 'B']
}
# 处理答题卡(假设有正反面)
result = recognizer.process_answer_sheet(
front_image='gaokao_answer_sheet_front.jpg',
back_image='gaokao_answer_sheet_back.jpg', # 可选
answer_key=answer_key
)
# 导出结果
if result['status'] == 'success':
recognizer.export_to_excel(result, 'recognition_result.xlsx')
原创声明:本文系作者授权腾讯云开发者社区发表,未经许可,不得转载。
如有侵权,请联系 cloudcommunity@tencent.com 删除。