关于验证码训练

开发调优
1

遇到一个验证码。如图：
image
image

想着自己生成一批，用ddddocr_trainer去训练一下。用cursor写了个代码，但是练出来后效果很差。有佬指点一下吗？

爬下来的验证码：

手动生成验证码的代码：

import random
from PIL import Image, ImageDraw, ImageFont

# 设置图像尺寸
width = 32
height = 24

# 创建新图像，黑底
image = Image.new("RGB", (width, height), "black")
draw = ImageDraw.Draw(image)

# 加载字体
font_size = 16  # 保持原字体大小，但噪点排除区域按8x8处理
font = ImageFont.truetype("font/typostuck.ttf", font_size)

# 生成4个随机字符（A-Z 和 1-9，不包含0）
chars = [random.choice("BCEFGHKMORTVXY2346789") for _ in range(4)]

# 设置字符位置和y偏移
x_positions = [0, 8, 17, 26]  # x位置：0, 8, 16, 24
y_offsets = [random.randint(-2, 2) for _ in range(4)]  # 随机y偏移

# 在图像上绘制字符（白字）
base_y = (height - 8) // 2  # 中心8x8区域，base_y = (24 - 8) // 2 = 8
for x, char, y_offset in zip(x_positions, chars, y_offsets):
    y = base_y + y_offset
    draw.text((x, y), char, font=font, fill="white")

# 定义噪点参数
noise_density = 0.5  # 噪点密度为80%
gap_x = {6, 7, 14, 15, 22, 23}  # 2像素间隙的x坐标

# 定义8x8噪点排除区域
noise_free_regions = []
for x_position, y_offset in zip(x_positions, y_offsets):
    x_start = x_position
    x_end = x_position + 10
    y_start = base_y + y_offset
    y_end = y_start + 10
    noise_free_regions.append({
        'x_range': (x_start, x_end),
        'y_range': (y_start, y_end)
    })

# 计算所有可放置噪点的像素
total_noise_pixels = 0
valid_pixels = []
for x in range(width):
    if x in gap_x:  # 排除2像素间隙
        continue
    for y in range(height):
        is_noise_free = False
        for region in noise_free_regions:
            x_start, x_end = region['x_range']
            y_start, y_end = region['y_range']
            if x_start <= x < x_end and y_start <= y < y_end:
                is_noise_free = True
                break
        if not is_noise_free:
            valid_pixels.append((x, y))
            total_noise_pixels += 1

# 计算噪点数量
num_noise_pixels = int(total_noise_pixels * noise_density)

# 添加噪点
for _ in range(num_noise_pixels):
    if valid_pixels:
        x, y = random.choice(valid_pixels)
        draw.point((x, y), fill="white")

# 保存图像
image.save("captcha.png")

ddddocr_trainer的训练参数：

Model:
    CharSet: [' ', '3', '6', K, V, '7', H, M, '5', B, F, O, Y, '4', '9', X, '2', C,
        G, E, '8', W, T, R]
    ImageChannel: 1
    ImageHeight: 64
    ImageWidth: -1
    Word: false
System:
    Allow_Ext: [jpg, jpeg, png, bmp]
    GPU: true
    GPU_ID: 0
    Path: captchas/
    Project: captcha_1m
    Val: 0.005
Train:
    BATCH_SIZE: 512
    CNN: {NAME: ddddocr}
    DROPOUT: 0.25
    LR: 0.002
    OPTIMIZER: Adam
    SAVE_CHECKPOINTS_STEP: 20000
    TARGET: {Accuracy: 0.999, Cost: 0.001, Epoch: 2}
    TEST_BATCH_SIZE: 512
    TEST_STEP: 10000
2

先对验证码进行一个简单的处理？比如只有单独的像素，就删除？

3

有道理。我先去试试