传统方法

程序

python
import numpy as np
import cv2
def replaceZeroes(data):
    min_nonzero = min(data[np.nonzero(data)])
    data[data == 0] = min_nonzero
    return data
def SSR(src_img, size):
    L_blur = cv2.GaussianBlur(src_img, (size, size), 0)
    img = replaceZeroes(src_img)
    L_blur = replaceZeroes(L_blur)
    dst_Img = cv2.log(img / 255.0)
    dst_Lblur = cv2.log(L_blur / 255.0)
    dst_IxL = cv2.multiply(dst_Img, dst_Lblur)
    log_R = cv2.subtract(dst_Img, dst_IxL)
    dst_R = cv2.normalize(log_R, None, 0, 255, cv2.NORM_MINMAX)
    log_uint8 = cv2.convertScaleAbs(dst_R)
    return log_uint8
# 使用照片修复的办法来去掉图片中的文字
def remove_text1(path):
    size = 3
    src_img = cv2.imread(path)
    # # 膨胀图片
    # src_img = cv2.dilate(src_img, kernel=np.ones((3, 3), dtype=np.uint8), iterations=2)
    # 腐蚀图片
    # src_img = cv2.erode(src_img, kernel=np.ones((3, 3), dtype=np.uint8), iterations=2)
    b_gray, g_gray, r_gray = cv2.split(src_img)
    b_gray = SSR(b_gray, size)
    g_gray = SSR(g_gray, size)
    r_gray = SSR(r_gray, size)
    ssrimg = cv2.merge([b_gray, g_gray, r_gray])
    canny_img = cv2.Canny(ssrimg, 200, 150)
    k = np.ones((3, 3), np.uint8)
    mask = cv2.morphologyEx(canny_img, cv2.MORPH_CLOSE, k)  # 闭运算
    # 对mask进行膨胀
    # mask = cv2.dilate(mask, kernel=np.ones((3, 3), dtype=np.uint8), iterations=6)
    dst = cv2.inpaint(src_img, mask, 3, cv2.INPAINT_TELEA)
    # dst = cv2.inpaint(dst, mask, 3, cv2.INPAINT_TELEA)
    # 膨胀
    dst = cv2.dilate(dst, kernel=np.ones((3, 3), dtype=np.uint8), iterations=5)
    return dst
# 使用mask直接填充颜色的办法来去掉图片中的文字
def remove_text2(path):
    size = 3
    src_img = cv2.imread(path)
    # 腐蚀图片
    # src_img = cv2.erode(src_img, kernel=np.ones((3, 3), dtype=np.uint8), iterations=2)
    b_gray, g_gray, r_gray = cv2.split(src_img)
    b_gray = SSR(b_gray, size)
    g_gray = SSR(g_gray, size)
    r_gray = SSR(r_gray, size)
    ssrimg = cv2.merge([b_gray, g_gray, r_gray])
    canny_img = cv2.Canny(ssrimg, 200, 150)
    k = np.ones((3, 3), np.uint8)
    mask = cv2.morphologyEx(canny_img, cv2.MORPH_CLOSE, k)  # 闭运算
    # 保存mask为图片
    cv2.imwrite("mask.png", mask)
    # mask等于1的区域，用白色填充
    # dst = cv2.bitwise_and(src_img, src_img, mask=mask)
    src_img[mask == 255] = (255, 255, 255)
    # 膨胀
    src_img = cv2.dilate(src_img, kernel=np.ones((3, 3), dtype=np.uint8), iterations=5)
    return src_img
# 移除背景色，增强文字
def remove_ground(path):
    size = 3
    src_img = cv2.imread(path)
    b_gray, g_gray, r_gray = cv2.split(src_img)
    b_gray = SSR(b_gray, size)
    g_gray = SSR(g_gray, size)
    r_gray = SSR(r_gray, size)
    ssrimg = cv2.merge([b_gray, g_gray, r_gray])
    canny_img = cv2.Canny(ssrimg, 200, 150)
    k = np.ones((3, 3), np.uint8)
    mask = cv2.morphologyEx(canny_img, cv2.MORPH_CLOSE, k)  # 闭运算
    # mask等于1的区域，用白色填充
    dst = cv2.bitwise_and(src_img, src_img, mask=mask)
    dst[mask == 0] = (255, 255, 255)
    return dst
if __name__ == '__main__':
    imgPath = "papertext.jpg"
    dst = remove_text2(imgPath)
    cv2.imwrite("result.jpg", dst)

原图：

找到文字部分：

去除文字：

保留清晰文字：

深度学习

1、

Self-Supervised Text Erasing with Controllable Image Synthesis 即STE ： 论文https://arxiv.org/abs/2204.12743。

2、

EnsNet ： https://blog.csdn.net/u013554402/article/details/104494261。

3、

https://paperswithcode.com/paper/mtrnet-one-stage-mask-based-scene-text-eraser。

4、

http://www.cjig.cn/html/jig/2022/4/20220420.htm

译文回填

SickZil-Machine用于去除文本： https://github.com/KUR-creative/SickZil-Machine

lama-cleaner技术： https://github.com/Sanster/lama-cleaner