为何Tesseract 4.1.1在清晰数字时钟图像上识别性能不佳？

我正在使用Tesseract 4.1.1开展数字时钟时间识别项目，但多数情况下无法识别内容，返回空结果。即便图像数据相当清晰，识别性能仍很差，明明Tesseract在更复杂任务上表现尚可，我已耗费大量时间却陷入瓶颈，恳请提供解决思路！

已尝试的优化措施 #

• 通过参数-c tessedit_char_whitelist=0123456789:.设置感兴趣字符白名单；
• 将图像转换为黑字白底的二值图像；
• 使用约120张图像、500轮迭代微调英文模型，增加迭代次数也未改善效果；
• 根据tessdoc建议，将图像缩放至字符x-height处于20-30像素区间。

实现代码 #

import pytesseract 
import math
import cv2
import numpy as np

def read_time(filename):
    # 白名单字符设置
    tesseract_config = '-c tessedit_char_whitelist=0123456789:.  --psm 6'  

    # 图像转二值图
    clock_image = cv2.imread(filename)
    gray_image = cv2.cvtColor(clock_image, cv2.COLOR_BGR2GRAY)
    _, binary_image = cv2.threshold(gray_image,200,255,cv2.THRESH_BINARY)

    # 判断背景是否为白色，否则反转图像
    is_white_background = False
    if not is_white_background:
        binary_image = cv2.bitwise_not(binary_image)

    # 自定义模型配置
    custom_model_config = '-c tessedit_char_whitelist=0123456789:. -l my_custom_model --psm 6'

    # 缩放图像至字符高度20-30像素区间
    cnt_black_pixels_per_row = np.sum(binary_image == 0, axis=1)
    non_zero_lines = [n for n,i in enumerate(cnt_black_pixels_per_row) if i > 0]

    # 获取字符上下边界估算当前高度
    top, bottom = non_zero_lines[0], non_zero_lines[-1]
    current_height = abs(top - bottom)
    height_expected = 32
    new_height = math.floor(binary_image.shape[0] * height_expected / current_height)
    new_shape = (binary_image.shape[1], new_height)
    final_image = cv2.resize(binary_image, new_shape, interpolation= cv2.INTER_LINEAR)

    # 四种方式尝试识别
    clock_text_binary = pytesseract.image_to_string(binary_image, config=tesseract_config)
    clock_text_binary_custom_config = pytesseract.image_to_string(binary_image, config=custom_model_config)
    clock_text_resized = pytesseract.image_to_string(final_image, config=custom_model_config)
    clock_text_resized_custom_config = pytesseract.image_to_string(final_image, config=tesseract_config)

    print(clock_text_binary, clock_text_binary_custom_config, clock_text_resized, clock_text_resized_custom_config)

识别失败的示例图像 #

内容的提问来源于stack exchange，提问作者h1t5uj1