CycleGAN生成车牌记录( 二 ) _gan

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
【CycleGAN生成车牌记录】训练集A图示例：
训练集B图示例：

文章插图
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-8import cv2import osimport numpy as npfrom PIL import ImageDraw, ImageFont, Imagechar_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 = provincenumber_list = np.random.choice(char_alphbet_list, 6)lp_str += "".join(number_list)lp_list.append(lp_str)return lp_listdef 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 *.ttcfont = 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_OpenCVdef generate_plate_rectangle(plate_str):"""绘制矩形边框:param plate_str::return:"""width = 256height = 256img = np.ones((height, width, 3), dtype=np.uint8)img *= 255# white backgroundcv2.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 imgdef generate_plate_filled_rectangle(plate_str):width = 256height = 256img = np.ones((height, width, 3), dtype=np.uint8)img *= 255# white backgroundcv2.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 imgdef 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手写汉字生成研究_参考网
对model稍作修改如下:
算法迭代了30+次后数据和效果如下：
算法迭代了130+次后数据和效果如下：
本次试验失败！！
四、只有文字倾斜的GAN
进一步查看GAN是否能够生成倾斜的字体，背景设置为黑色。