使用如下脚本切割获取了200多手写字体（图1、图2），249个字体训练后生成的字体（图3）和手写字体差别很大，跟手写字体文件名是否有关？

[liusenb]

使用如下脚本切割获取了200多手写字体（图1、图2），249个字体训练后生成的字体（图3）和手写字体差别很大，跟手写字体文件名是否有关？

import argparse
import os
import cv2
import shutil

def main(opt):
    # 膨胀腐蚀大小 即将笔画膨胀 避免将偏旁单独分割 根据图片请况自行设置
    rect_size = 16
    # 字体小于该值忽略 20*20
    ignore_min_size = 20
    # 字体大于该值忽略 100*100
    ignore_max_size = 100

    # 需要切分的图片 input/input.jpg
    input_file = opt.input_path

    # 输出文件夹 output
    output_path = opt.output_path + os.path.sep
    if not os.path.exists(output_path):
        os.makedirs(output_path)
    else:
        shutil.rmtree(output_path)
        os.makedirs(output_path)

    img = cv2.imread(input_file)

    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    # 二值化 根据图片情况调节灰度阈值 第二个参数灰度值表示小于该值就将其变为0
    ret, thresh = cv2.threshold(gray, 150, 255, cv2.THRESH_BINARY)
    cv2.imwrite(output_path + "thresh.jpg", thresh)
    thresh_rgb = cv2.cvtColor(thresh, cv2.COLOR_GRAY2BGR)
    # 待切分的图片 img原图 gray灰度图 thresh二值化后的灰度图 thresh_rgb二值化后转RGB
    result = thresh.copy()

    # 膨胀腐蚀 将字体笔画扩大15像素
    kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (rect_size, rect_size))
    eroded = cv2.erode(thresh, kernel)
    cv2.imwrite(output_path + "eroded.jpg", eroded)

    # 轮廓检测
    contours, hierarchy = cv2.findContours(eroded, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)

    # 膨胀间距
    spacing = rect_size // 2 - 1
    cut_frame_color = (0, 255, 0)
    font_frame_color = (255, 0, 0)
    for c in contours:
        x, y, w, h = cv2.boundingRect(c)
        if w < ignore_min_size or h < ignore_min_size or w > ignore_max_size or h > ignore_max_size:
            continue

        font_w = w - spacing - spacing
        font_h = h - spacing - spacing

        frame_size = font_w if font_w > font_h else font_h

        # 正方形切割偏移量
        x_offset = int((frame_size - font_w) / 2)
        y_offset = int((frame_size - font_h) / 2)

        start_x = x + spacing - x_offset
        start_y = y + spacing - y_offset
        end_x = x + w - spacing + x_offset
        end_y = y + h - spacing + y_offset

        # 字体框线
        cv2.rectangle(thresh_rgb, (x + spacing, y + spacing), (x + w - spacing, y + h - spacing), font_frame_color, 1)
        # 切割框线
        cv2.rectangle(thresh_rgb, (start_x, start_y), (end_x, end_y), cut_frame_color, 1)
        # 切割
        temp = result[start_y:end_y, start_x:end_x]
        path = opt.output_path + os.path.sep + "result" + os.path.sep
        if not os.path.exists(path):
            os.makedirs(path)
        cv2.imwrite(path + str(x) + "_" + str(y) + ".jpg", temp)

    cv2.imwrite(output_path + "result.jpg", thresh_rgb)

if __name__ == '__main__':
    """Parse input arguments"""
    parser = argparse.ArgumentParser()
    parser.add_argument('--input', dest='input_path', default='input/input.jpg',
                        help='Please set the input path')
    parser.add_argument('--output', dest='output_path', default='output',
                        help='Please set the output path')
    opt = parser.parse_args()
    main(opt)

Originally posted by @liusenb in https://github.com/dailenson/SDT/issues/43#issuecomment-1925431110

[liusenb]

使用user_generate.py脚本生成

[dailenson]

跟文件名应该没关系，可以试试对裁剪出来的图片做个简单的二值化处理，应该能提升效果https://github.com/dailenson/SDT/issues/59#issuecomment-1963233481