Question

我正在尝试击败程序中的反机器人功能，该程序中用户必须按字母数字顺序单击图像中的字母。我已经设法使用预处理从背景中提取了一些文本，但是仍然需要对每个单个字符进行去歪斜才能使用Tesseract获得最佳效果。

使用霍夫线之前的图像，仅进行预处理

绘制了由HoughLinesP检测到的线条的原始图像我尝试使用Canny Edge Detector + Hough Lines尝试查找每个字符下方的行。但是，它认为行的质量不一致并且过于依赖行，因此我无法区分底线和字符本身上检测到的行。

这是我尝试过的代码：

# -*- coding:utf-8 -*-
import cv2, numpy as np, time
img_roi = [48, 191, 980, 656]  # x1, y1, x2, y2
src_img_dir = "images/source/9.png"
bg_img = cv2.imread("images/background.png", cv2.IMREAD_COLOR)[img_roi[1]:img_roi[3], img_roi[0]:img_roi[2]]
# The background of the area is constant. So I have used a reference background image and removed pixels which have a similar H value as the background

bg_hsv = cv2.cvtColor(bg_img, cv2.COLOR_BGR2HSV)
src_img = cv2.imread(src_img_dir, cv2.IMREAD_COLOR)[img_roi[1]:img_roi[3], img_roi[0]:img_roi[2]]
# This image is the image where letters are placed on top of the background image

src_hsv = cv2.cvtColor(src_img, cv2.COLOR_BGR2HSV)
mask = np.zeros([src_img.shape[0], src_img.shape[1], 3], dtype=np.uint8)

offset = 3
start_time = time.time()
for y in range(src_img.shape[0]):
    for x in range(src_img.shape[1]):
        sp = src_hsv[y][x]
        bp = bg_hsv[y][x]

        if bp[0]-offset <= sp[0] <= bp[0]+offset:
            if sp[1] >= 109:
                mask[y][x] = src_img[y][x]
        elif sp[1] <= 90:
            if sp[0] >= 67:
                mask[y][x] = src_img[y][x]
            elif sp[2] >= 125 and sp[1] >= 20:
                mask[y][x] = src_img[y][x]
        else:
            mask[y][x] = src_img[y][x]
        """if sp[1] >= 60 and sp[2] >= 60:
            mask[y][x] = src_img[y][x]
            #mask[y][x] = conv"""

print("duration", time.time()-start_time)
mask = cv2.cvtColor(mask, cv2.COLOR_BGR2HSV)
#mask[:,:,2] = 255
mask = cv2.cvtColor(mask, cv2.COLOR_HSV2BGR)
mask_gray = cv2.cvtColor(mask, cv2.COLOR_BGR2GRAY)
ret, thresh = cv2.threshold(mask_gray, 0, 255, cv2.THRESH_BINARY | cv2.THRESH_OTSU)
opened = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, (3,3))
opened = cv2.morphologyEx(opened, cv2.MORPH_OPEN, (3,3))
opened = cv2.erode(opened, (3,3))
opened = cv2.dilate(opened, (3,3))
opened = cv2.dilate(opened, (5, 5))
opened = cv2.morphologyEx(opened, cv2.MORPH_CLOSE, (3,3))
opened = cv2.erode(opened, (3,3))
opened = cv2.erode(opened, (3,3))
final_img = opened
#edges = cv2.Canny(final_img, 0, 255)
lines = cv2.HoughLinesP(final_img, 1, np.pi / 180, 20, minLineLength=10, maxLineGap=3)
for line in lines:
        coords = line[0]
        cv2.line(src_img, (coords[0], coords[1]), (coords[2], coords[3]), [255,255,255], 2)
#cv2.imshow("can", edges)


#cv2.drawContours(src_img, fixed_contours, -1, (0,255,0), 2)
cv2.imshow("src", src_img)
cv2.imshow("", final_img)

cv2.waitKey(0)
cv2.destroyAllWindows()

Answer 1

乍一看，似乎偏斜不是很强，而且字符相距很远。

我会对经过过滤的图像进行多步处理（已经相当不错了）

首先检测包含两个明显较大斑点（字母/数字+下划线）的感兴趣区域，过滤出噪声像素
然后将下划线明确检测为下划线和下划线两者的长而平坦（在这方面，字母“ I”和数字“ 1”可能会出现问题）
根据感兴趣的局部区域（下划线+字符）使用下划线方向来确定哪一侧朝下
试探性地确定偏斜角：假设x度（在x的狭窄范围内循环），则感兴趣的局部区域的信号在下划线上方的四边形内，使得底部（下划线）与左侧之间的角度为x。 li>
使用图像不变形功能，以便将下划线映射到宽高比适当的矩形的底边
利润

对图像中的单个字符进行倾斜校正

1 个答案: