Research On Recognition Method Of Paper Currency Serial Number For Transmission Imaging

There are two imaging methods for paper currency. One is the catoptric imaging.The other is the transmission imaging. RMB images of catoptric imaging are clear and itsbackground is simple. But two imaging sensors are necessary to capture images of papercurrency that are put into equipments in arbitrary direction. Large room is needed for towimaging sensors. While only one imaging sensor is needed for transmission imaging insame conditions. Transmission imaging has the advantage of small necessary room. Nowseveral hundred thousand ATMs need to update by adding serial number recognitionfunction. Due to the space limitation of its inhere design, transmission imaging has to beadopted, which saves room. But images of transmission imaging have complexbackgrounds, double-sided overlap and prone to uneven brightness. All above bring hugedifficulty to the location of serial numbers.To solve problems appeared in paper money serial numbers images of transmissionimaging, a whole recognition plan for transmission imaging is proposed. This planincludes location of serial numbers region, location of single character, imagepreprocessing and character recognition. A new algorithm for location of serial numbersregion based on average gray value of sub-region is proposed. Firstly, detect featurepoints of margin between characters. Namely, construct a slidable window and divide thewindow into several sub-regions. Label candidate serial numbers’ region with a specialmark by comparing different grayscale average value among different sub-regions. Then horizontal location of serial numbers region is done. Horizontal projection is done on thefoundation of rough location. And then calculate the accurate horizontal location of serialnumbers region based on inherent attributes of RMB. After that, calculate the accuratevertical location indirectly by locating Omron rings on the foundation of the horizontallocation. Finally locate single characters by exploiting connected component labeling.For potential cracked and adhesive characters, split and merge are done specially. Noisesare eliminated in the image after region label through some attributes, such as area andlength-width ration, with respect to image preprocessing. A new method using templatematching to recognize first three characters and structural method to recognize last sevencharacters is proposed to eliminate influence of transmission imaging with respect tocharacter recognition.