CycleGAN生成车牌记录

数据猴赛雷

已于 2023-03-27 10:14:50 修改

阅读量632

点赞数 1

文章标签： opencv 计算机视觉人工智能

于 2023-03-24 09:55:41 首次发布

本文链接：https://blog.csdn.net/benben044/article/details/129724149

版权

本次代码开发调试基于简化版的CycleGAN，相关代码说明见文档：

CycleGAN二次解读_benben044的博客-CSDN博客

一、直接使用高仿车牌生成车牌

高仿车牌生成的代码见：python+opencv生成较真实的车牌号码图片_benben044的博客-CSDN博客

从网上下载了一批真实的质量较高的车牌，每张车牌都有车牌号码的文字信息，如下图所示：

首先，根据上图中的车牌号码生成高仿的车牌，作为训练集的A目录。

然后，将真实车牌作为训练集的B目录。

我们将上述训练集通过cycleGAN进行训练，200次迭代后，效果依然不佳。

效果大致如下：

然后将迭代次数提高到400次，以及增加ResnetBlock的个数，效果依然不佳。

所以提出了疑问：

1、cycleGAN的示例中的图片（比如苹果<->橘子）是不涉及文字的，那么这种涉及文字的是否可行？

2、自己训练集的图片是矩形的，而代码中会resize为正方形，这块是否有影响。

带着这两个疑问，我们重新设计了训练集进行测试。

二、最简单的带文字的GAN试验

cycleGAN源码中图片最后会被resize到256*256，所以这次我们直接生成256*256的图片。

源图片和目标图片样式分别见下面两张图：

生成代码：

#!/usr/bin/env python
#coding=utf-8

import cv2
import os
import numpy as np
from PIL import ImageDraw, ImageFont, Image

char_province_list = ["京","沪","津","渝","冀","晋","蒙","辽","吉","黑","苏","浙","皖","闽","赣","鲁","豫","鄂","湘","粤","桂","琼","川","贵","云","藏","陕","甘","青","宁","新"]
char_alphbet_list = ["0","1","2","3","4","5","6","7","8","9","A","B","C","D","E","F","G","H","J","K","L","M","N","P","Q","R","S","T","U","V","W","X","Y","Z"]

def generate_license_plate_number():
    """
    随机生成车牌号码
    :return:
    """
    lp_list = []
    for province in char_province_list:
        for _ in range(100):
            lp_str = province
            number_list = np.random.choice(char_alphbet_list, 6)
            lp_str += "".join(number_list)
            lp_list.append(lp_str)
    return lp_list

def write_chinese(img, font_type, font_size,color, position, content):
    # 图像从OpenCV格式转换成PIL格式
    img_PIL = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
    # 字体  字体*.ttc的存放路径一般是： /usr/share/fonts/opentype/noto/ 查找指令locate *.ttc
    font = ImageFont.truetype(font_type, font_size)
    # 字体颜色
    # 文字输出位置
    # 输出内容
    draw = ImageDraw.Draw(img_PIL)
    draw.text(position, content, font=font, fill=color)
    # 转换回OpenCV格式
    img_OpenCV = cv2.cvtColor(np.asarray(img_PIL), cv2.COLOR_RGB2BGR)
    return img_OpenCV

def generate_plate_rectangle(plate_str):
    """
    绘制矩形边框
    :param plate_str:
    :return:
    """
    width = 256
    height = 256
    img = np.ones((height, width, 3), dtype=np.uint8)
    img *= 255  # white background
    cv2.rectangle(img, (20, 20), (236, 236), (255, 0, 0), 4)
    img = write_chinese(img, 'font/SimHei.ttf', 42, (0, 0, 0), (50, 100), plate_str)
    return img

def generate_plate_filled_rectangle(plate_str):
    width = 256
    height = 256
    img = np.ones((height, width, 3), dtype=np.uint8)
    img *= 255  # white background
    cv2.rectangle(img, (20, 20), (236, 236), (255, 0, 0), -1)
    img = write_chinese(img, 'font/SimHei.ttf', 42, (0, 0, 0), (50, 100), plate_str)
    return img


def generate_plate_tuple(plate_str, index):
    dir_path = "D:\\workspace\\ncz-python-algo\\com\\ncz\\algo\\license-plate-generator\\simple_images"
    path_A = os.path.join(dir_path, "A", str(index) + "_" + plate_str + ".png")
    path_B = os.path.join(dir_path, "B", str(index) + "_" + plate_str + ".png")

    img1 = generate_plate_rectangle(plate_str)
    img2 = generate_plate_filled_rectangle(plate_str)
    cv2.imencode('.png', img1)[1].tofile(path_A)
    cv2.imencode('.png', img2)[1].tofile(path_B)


# # 生成个别图片
# license_plate_str = '浙A5B5T3'
# generate_plate_tuple(license_plate_str, 1)

# 随机生成一堆图片
license_plate_list = generate_license_plate_number()
for index, license_plate in enumerate(license_plate_list):
    print(index + 1, license_plate)
    generate_plate_tuple(license_plate, index + 1)

经过200次迭代训练之后，对应的loss信息如下：

将模型进行测试效果如下（本车牌未在训练集中）：

B生成A的图片：

A生成B的图片：

可以看到，GAN生成的图片除了少部分瑕疵，基本上是符合预期的，同时文字是清晰可见的，没有出现第一部分试验所看到的文字错乱的情况。

针对上述好的迹象，我们再进一步一点点，看下B图中文字倾斜的话，文字是否可以正确生成。

三、倾斜的带文字的GAN试验

训练集A图示例：

训练集B图示例：

生成训练集代码（在第二步的基础上稍作修改）：

#!/usr/bin/env python
#coding=utf-8

import cv2
import os
import numpy as np
from PIL import ImageDraw, ImageFont, Image

char_province_list = ["京","沪","津","渝","冀","晋","蒙","辽","吉","黑","苏","浙","皖","闽","赣","鲁","豫","鄂","湘","粤","桂","琼","川","贵","云","藏","陕","甘","青","宁","新"]
char_alphbet_list = ["0","1","2","3","4","5","6","7","8","9","A","B","C","D","E","F","G","H","J","K","L","M","N","P","Q","R","S","T","U","V","W","X","Y","Z"]

def generate_license_plate_number():
    """
    随机生成车牌号码
    :return:
    """
    lp_list = []
    for province in char_province_list:
        for _ in range(100):
            lp_str = province
            number_list = np.random.choice(char_alphbet_list, 6)
            lp_str += "".join(number_list)
            lp_list.append(lp_str)
    return lp_list

def write_chinese(img, font_type, font_size,color, position, content, degree):
    # 图像从OpenCV格式转换成PIL格式
    img_PIL = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
    # 字体  字体*.ttc的存放路径一般是： /usr/share/fonts/opentype/noto/ 查找指令locate *.ttc
    font = ImageFont.truetype(font_type, font_size)
    # 字体颜色
    # 文字输出位置
    # 输出内容
    draw = ImageDraw.Draw(img_PIL)
    draw.text(position, content, font=font, fill=color)
    img_PIL = img_PIL.rotate(degree)
    # 转换回OpenCV格式
    img_OpenCV = cv2.cvtColor(np.asarray(img_PIL), cv2.COLOR_RGB2BGR)
    return img_OpenCV

def generate_plate_rectangle(plate_str):
    """
    绘制矩形边框
    :param plate_str:
    :return:
    """
    width = 256
    height = 256
    img = np.ones((height, width, 3), dtype=np.uint8)
    img *= 255  # white background
    cv2.rectangle(img, (20, 20), (236, 236), (255, 0, 0), 4)
    img = write_chinese(img, 'font/SimHei.ttf', 42, (0, 0, 0), (50, 100), plate_str, 0)
    return img

def generate_plate_filled_rectangle(plate_str):
    width = 256
    height = 256
    img = np.ones((height, width, 3), dtype=np.uint8)
    img *= 255  # white background
    cv2.rectangle(img, (20, 20), (236, 236), (255, 0, 0), -1)
    img = write_chinese(img, 'font/SimHei.ttf', 42, (0, 0, 0), (50, 100), plate_str, 20)
    return img


def generate_plate_tuple(plate_str, index):
    dir_path = "D:\\workspace\\ncz-python-algo\\com\\ncz\\algo\\license-plate-generator\\simple_images"
    path_A = os.path.join(dir_path, "A", str(index) + "_" + plate_str + ".png")
    path_B = os.path.join(dir_path, "B", str(index) + "_" + plate_str + ".png")

    img1 = generate_plate_rectangle(plate_str)
    img2 = generate_plate_filled_rectangle(plate_str)
    cv2.imencode('.png', img1)[1].tofile(path_A)
    cv2.imencode('.png', img2)[1].tofile(path_B)


# 生成个别图片
# license_plate_str = '浙A5B5T3'
# generate_plate_tuple(license_plate_str, 1)

#随机生成一堆图片
license_plate_list = generate_license_plate_number()
for index, license_plate in enumerate(license_plate_list):
    print(index + 1, license_plate)
    generate_plate_tuple(license_plate, index + 1)

参考文档：基于亚像素卷积的改进型CycleGAN手写汉字生成研究_参考网

对model稍作修改如下: