Tesseract数字识别不准确，如何优化代码提升准确率？

问题背景 #

有一批仅包含数字的图片，已完成预处理但Tesseract仍无法准确提取数字，典型错误案例：

• 识别结果0858，预期结果08588（漏识别数字）
• 识别结果95415，预期结果92412（数字混淆）
• 识别结果2043，预期结果20413（漏识别数字）
• 识别结果61416，预期结果61116（数字混淆）

原实现代码：

from PIL import Image, ImageEnhance, ImageFilter
import pytesseract


CAPTCHA_PATH = 'captcha_images/captcha8.jpeg'
RED_REMOVED_PATH = 'processed/red_removed.jpeg'
PROCESSED_IMAGE_PATH = 'processed/processed_image.png'


def change_pixels_except_black(image_path, output_path, threshold=50):
    """
    Change all pixels to white except for pixels close to black.

    :param image_path: Path to the input image.
    :param output_path: Path to save the output image.
    :param threshold: Threshold to determine if a pixel is black (default is 50).
    """
    # Open the image
    image = Image.open(image_path)
    image = image.convert("RGB")  # Ensure the image is in RGB mode

    # Load the image data
    pixels = image.load()

    # Get the dimensions of the image
    width, height = image.size

    # Iterate over each pixel
    for y in range(height):
        for x in range(width):
            # Get the current pixel's color
            r, g, b = pixels[x, y]

            # Check if the pixel is close to black
            if r < threshold and g < threshold and b < threshold:
                # Keep the pixel as is (close to black)
                continue
            else:
                # Change the pixel to white
                pixels[x, y] = (255, 255, 255)

    # Save the modified image
    image.save(output_path)


def preprocess_image(image_path, output_path):
    """
    Preprocess the image to enhance its quality for OCR.

    :param image_path: Path to the input image.
    :param output_path: Path to save the processed image.
    """
    
    # Open the image
    image = Image.open(image_path)

    # Resize the image
    new_width = image.width * 2
    new_height = image.height * 2
    image = image.resize((new_width, new_height), Image.LANCZOS)

    # Convert to grayscale
    image = image.convert('L')

    # Increase contrast
    enhancer = ImageEnhance.Contrast(image)
    image = enhancer.enhance(2)

    # Apply a filter to sharpen the image
    image = image.filter(ImageFilter.SHARPEN)

    # Save the processed image
    image.save(output_path)


def extract_text(image_path):
    """
    Extract text from an image using pytesseract.

    :param image_path: Path to the input image.
    :return: Extracted text.
    """
    text = pytesseract.image_to_string(Image.open(
        image_path), lang='eng', config='--psm 10 --oem 3 -c tessedit_char_whitelist=0123456789')

    return text


# Example usage
change_pixels_except_black(CAPTCHA_PATH, RED_REMOVED_PATH, threshold=70)
preprocess_image(RED_REMOVED_PATH, PROCESSED_IMAGE_PATH)
text = extract_text(PROCESSED_IMAGE_PATH)
print(f"Extracted Text of {CAPTCHA_PATH}:")
print(text)

改进方案 #

1. 替换对比度增强为二值化处理 #

当前的对比度增强无法彻底分离数字与背景，改用阈值二值化让数字边缘更锐利，减少模糊干扰：

def preprocess_image(image_path, output_path):
    image = Image.open(image_path)
    # 放大图片提升分辨率
    new_width = image.width * 2
    new_height = image.height * 2
    image = image.resize((new_width, new_height), Image.LANCZOS)
    # 转灰度图
    image = image.convert('L')
    # 自适应阈值二值化（适合光照不均的图片）
    image = image.point(lambda x: 0 if x < 127 else 255, '1')
    # 若图片光照均匀，可改用固定阈值（根据实际调整阈值）
    # image = image.point(lambda x: 0 if x < 150 else 255, '1')
    # 二次锐化强化笔画
    image = image.filter(ImageFilter.SHARPEN)
    image.save(output_path)

2. 调整Tesseract的PSM模式 #

当前使用--psm 10（单字符识别模式），但图片是多数字序列，应选用适合单行文本的模式：

def extract_text(image_path):
    # --psm 7：将图片视为单行文本；若确定是固定长度数字，可试--psm 6（单块文本）
    text = pytesseract.image_to_string(Image.open(image_path), 
                                      lang='eng', 
                                      config='--psm 7 --oem 3 -c tessedit_char_whitelist=0123456789')
    # 清理识别结果中的空格、换行符
    text = text.strip()
    return text

3. 引入OpenCV做形态学降噪 #

如果图片存在细小噪点或笔画残缺，用OpenCV的形态学操作优化预处理效果：

import cv2
import numpy as np

def preprocess_with_opencv(image_path, output_path):
    # 读取灰度图
    img = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
    # 放大图片
    img = cv2.resize(img, None, fx=2, fy=2, interpolation=cv2.INTER_LANCZOS4)
    # 自适应阈值二值化（反转黑白，让数字为白色、背景为黑色）
    img = cv2.adaptiveThreshold(img, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 11, 2)
    # 去除微小噪点
    kernel = np.ones((1,1), np.uint8)
    img = cv2.morphologyEx(img, cv2.MORPH_OPEN, kernel)
    # 保存处理后的图片
    cv2.imwrite(output_path, img)

使用时替换原有的preprocess_image函数即可，OpenCV对验证码类图片的处理效果优于单纯的PIL操作。

4. 针对数字混淆的针对性优化 #

对于2/5、1/4这类易混淆数字：

• 微调二值化阈值，确保数字笔画完整无断裂
• 若数字长度固定，可在识别后做长度校验，比如预期5位但识别出4位，可检查预处理后的图片是否有漏识别的区域
• 训练Tesseract自定义数字模型：准备标注好的数字样本，生成专属的.traineddata文件，提升特定字体数字的识别准确率

5. 验证预处理效果 #

每次调整预处理步骤后，务必查看生成的processed_image.png——如果肉眼都难以分辨数字，Tesseract不可能识别准确，优先保证预处理后的图片数字清晰、无多余噪点、笔画连贯。

内容的提问来源于stack exchange，提问作者Vishwa Mittar