OCR识别中,某些图像中的字符和背景之间的亮暗差别很小,需要进行处理,将字符区域从图像中扣出来进行识别。
*
* This example describes one step from the semiconductor product chain.
* In the front-end-of-line step, the ICs are printed on the wafer. To
* tag a single wafer from the production life line, each wafer receives
* an ID number, printed with the SEMI font. This ID number is read here.
*
* 读图像
dev_update_off ()
dev_close_window ()
read_image (Image, 'ocr/wafer_semi_font_01')
dev_open_window_fit_image (Image, 0, 0, -1, -1, WindowHandle)
dev_set_draw ('margin')
set_display_font (WindowHandle, 16, 'mono', 'true', 'false')
dev_set_line_width (2)
dev_set_colored (12)
*
* 读指定分类器
read_ocr_class_mlp ('SEMI', OCRHandle)
NumImages := 10
for Index := 1 to NumImages by 1
*
* Segment characters
* 读入一张图像
read_image (Image, 'ocr/wafer_semi_font_' + Index$'02')
* Characters must be black-on-white, i.e., dark characters on a light background
* 翻转图像,目前是黑底白字,而字符识别,需要的是白底黑字
* 此翻转将图像变为白底黑字
invert_image (Image, ImageInvert)
* 下面两个算子一般一块使用
* 对于mean_image的参数mask的大小,一般取需要的目标的大小的2倍;如目标是个圆形,直径为10,则mask的大小一般取20
* 对于一些目标与背景的阈值对比不明显的图像,需要用到动态阈值分割,使用局部阈值来分割图像
* 均值滤波
* 对于背景和前景之间差别不大的图像,可以使用如下两个算子进行处理
mean_image (Image, ImageMean, 31, 31)
* 阈值处理,ImageMean一般为通过均值滤波处理之后的区域
dyn_threshold (Image, ImageMean, RegionDynThresh, 7, 'light')
* Characters are often dotted. Therefore, we first merge close dots
* that belong to the same character just before calling the operator connection
*闭运算,让一个字符能成为一个连通域,然后再进行connection计算,得到单个字符的连通域
closing_circle (RegionDynThresh, RegionClosing, 2.0)
connection (RegionClosing, ConnectedRegions)
* Filter out characters based on two facts:
* 1. Characters are printed in SEMI-12. Therefore we can make strong assumptions
* on the dimensions of the characters
* 2. Characters are printed along a straight line
* 通过特征直方图,将字符的区域提取出来,
* 第一步提取使用的特征是宽和高
* 第二步提取使用的特征是行坐标
select_shape (ConnectedRegions, SelectedRegions1, ['height','width'], 'and', [29,15], [60,40])
area_center (SelectedRegions1, Area, RowCh, ColumnCh)
MedianRow := median(RowCh)
select_shape (SelectedRegions1, Chars, 'row', 'and', MedianRow - 30, MedianRow + 30)
*
* Read out segmented characters
* 字符区域排序,同时获得各个字符的方形区域
sort_region (Chars, CharsSorted, 'character', 'true', 'column')
* 将字符区域转成矩形区域
shape_trans (CharsSorted, Characters, 'rectangle1')
* 字符识别
do_ocr_multi_class_mlp (Characters, ImageInvert, OCRHandle, Class, Confidence)
*
* 显示识别出来的字符
dev_display (ImageInvert)
dilation_rectangle1 (Characters, RegionDilation, 7, 7)
dev_display (RegionDilation)
area_center (CharsSorted, Area1, Row, Column)
MeanRow := mean(Row)
for IndexL := 0 to |Class| - 1 by 1
disp_message (WindowHandle, Class[IndexL], 'image', MeanRow + 40, Column[IndexL] - 20, 'black', 'true')
endfor
if (Index != NumImages)
disp_continue_message (WindowHandle, 'black', 'true')
stop ()
endif
endfor
* 释放分类器资源
clear_ocr_class_mlp (OCRHandle)
进行OCR识别的图像,如下图所示:
可以看到,这种图像的字符和背景之间的差别很小,我们需要将字符区域从这样的图像中提取出来。
在对图像进行预处理和Blob分析中,针对此类图像,有如下两个算子是比较关键的:
* 对于背景和前景之间差别不大的图像,可以使用如下两个算子进行处理
mean_image (Image, ImageMean, 31, 31)
* 阈值处理,ImageMean一般为通过均值滤波处理之后的区域
dyn_threshold (Image, ImageMean, RegionDynThresh, 7, 'light')
通过预处理和Blob分析,可以得到如下字符区域:
将以上字符区域放入字符识别算子,进行识别,得到的结果如下:
更多halcon代码分析,见如下公众号: