Cigarette's Unique Code Detection And Recognition Based On Deep Learning

Cigarette's unique code plays a very important role in tobacco industry.A Cigarette's unique code for a cigarette is like an ID card for a citizen.It is the identity of a cigarette.The process of checking 2-line 32-bit cigarette's unique code requires staff to check one by one,which needs a lot of manpower and energy,which is also prone to errors.This paper proposes an OCR(Optical Character Recognition)algorithm based on deep learning,which can quickly and automatically recognize the target of cigarette code by taking photos.As a developed technology,OCR has been used in many fields.However,due to the complexity and diversity of cigarette case itself,the existing OCR software on the market can not work well in this field.Also,deep learning models need to obtain a large num-ber of real data of cigarette code,which requires a lot of manpower;in addition,because the background and target characters are difficult to distinguish,locating the 32 characters becomes more difficult.To solve the above problems,the algorithm is divided into two parts,location(detection)and recognition.In the recognition part,the data set is expanded by means of artificial data enhancement,and the recognition model based on CRNN neural network structure is im-proved;in the detection part,the two-direction sliding windows from top to bottom and from bottom to top are used to scan and analyze the images,and locate the position of the characters.In the experiment,we focus on reducing the time consumption and optimizing the accuracy.Finally,we developed a cigarette's code recognition system with locating accuracy rate of more than 93%,16-character recognition accuracy rate of 70%,and single character recognition accuracy rate of 98%.The process of recognizing one image only costs less than 0.5 second.