在深度学习中，样本不均衡是指不同类别的样本数据量差别比较大，利用不均衡样本训练出来的模型泛化能力差，并且容易发生过拟合。

对于图像的分类问题，通过各个类别的样本数量就可以判断出训练样本是否过拟合。对于OCR识别问题，我们的目的是准确的识别出文本中的每个字符，训练样本通常是不定长度不同字符的组合，因此这里的样本不均衡指的是不同字符的数量差别比较大，无法简单通过图像的数量判断，但是可以对训练样本和对应的标注文档进行遍历，从而获得字典中每个字符的出现频率，进一步判断训练样本是否均衡！
贴上完整的代码！

最后贴上完整的代码：

# -*- coding: utf-8 -*-

# Usage:
# python /input/image/and/txt/folder/path/ /output/folder/path

import os
import sys
import glob as gb

def get_label_files(folder):
    in_path = os.path.join(folder, "*.txt")
    files = []
    for txt_file in gb.glob(in_path):
        img_file = txt_file[:-3] + "jpg"
        if os.path.exists(img_file):
            files.append(os.path.basename(txt_file))
    return files


if __name__ == '__main__':

    # input_folder = sys.argv[1]
    # output_folder = sys.argv[2]

    # get images and corresponding txt files
    files = []
    files = get_label_files(input_folder)

    # get dict
    dict_file = './dml_digital.txt'
    result = dict()
    with open(dict_file, 'r') as file:
        for line in file:
            line = line.decode('utf-8').strip()
            result[line] = 0

    # trace files to get statistics
    for label_file in files:
        with open(os.path.join(input_folder, label_file)) as file:
            text = file.read().decode("utf-8").strip()
            length = len(text)
            for k in range(length):
                if result.has_key(text[k]):
                    result[text[k]] += 1

    # write statistics
    output_file_path = os.path.join(output_folder,os.path.basename(input_folder) + '.txt')

    with open(output_file_path, 'w+') as f:
        for key, val in result.items():
            f.write(key.encode("utf-8") + ' ' + str(val) + '\n')
    print('Finished calculation!')

统计结果.jpeg

其实样本的均衡性判断是训练前的工作，活生生的被我拖到了现在，又一次战胜了拖延症，奖励自己一块西瓜🍉！