Problem in recognition of a complex shape as one single character!

[mhhma...@gmail.com]

Dear All,
I am using tesseract_ocr in order to train and recognize some Perisan characters plus some shapes. In recognition of all the characters and shapes, tesseract works well except of one shape that has many children and is more complex than others.
for example, my trained tesseract, recognize the following characters well with high confidence:






the confidence for the above characters and shapes is higher than 85 and the API WordFontAttributes pointsize parameter is 12 for all of them. But for the following shape (the only shape that my trained tesseract doesn't work well), I catch the pointsize parameter as 4 and I have very low confidence about 35%. 



For solving this problem, I used the API parameters of tesseract proposed for Chinese and Japanese language in the group as:

tesseract.SetVariable("save_blob_choices", "T");
tesseract.SetVariable("chop_enable","T");
tesseract.SetVariable("use_new_state_cost","F");
tesseract.SetVariable("segment_segcost_rating","F");
tesseract.SetVariable("enable_new_segsearch","0");
tesseract.SetVariable("language_model_ngram_on","0");
tesseract.SetVariable("textord_force_make_prop_words","F");
tesseract.SetVariable("edges_max_children_per_outline","40");//end of our part!
tesseract.SetPageSegMode(PSM_SINGLE_CHAR);

but I have no change in my results.
It should be mentioned that if we don't use tesseract.SetPageSegMode(PSM_SINGLE_CHAR), tesseract consider the complex shape as 4 - 5 different characters.
I would greatly appreciate for any help.
Regard,
Hasan

[mhhma...@gmail.com]

Today I found an API parameter of tesseract that can solve my problem. But I don't know the exact reason!!!
By changing "textord_max_noise_size" (for example from 7 to 15), the confidence of recognition of my complex shape has been increased to more than 85 percent and this change, has not decrease the recognition confidence the other characters.
I don't understand exactly why this change in "textord_max_noise_size" can improve the accuracy and besides of this improvement, what we miss as a trade off?
Any Idea?
Regars, Hasan