Bank Notes Handwritten Digit String Recognition Preprocessing And Segmentation

Notes handwritten digits recognition system is of great practical value for banking industry. To the scanned images of bills, the main processing steps include two parts:character recognition, preprocessing and segmentation. From the existing recognition algorithms, we can see that the correct rate of recognition of a single digital character is higher than 99%. The key to improve the digit string recognition rate is to digital image preprocessing and segmentation.In this paper, we consider the handwritten numeral strings of bank check from one bank of construction as the background. We did some research and improvement on the core technology of digital image preprocessing and segmentation. The main contents include:1. Change color image to gray image. Considered the red seal in the image, we raised an algorithm to remove the effect of the seal, and retain the information of characters of numeral strings. In the processing, we enhanced the image and removed the noice, which made fondation for the subsequent work.2. Remove the frame of numeral strings, which contains two parts:the detection of frame line and the detection of intersection between frame line and character. Sometimes, the frame line is tilt and curved, we first detect the straight line, then find the exact edge points, finally picewise linear fitting method to accurately describe the frame line. In the processing of detection of intersection, we raised an algorithm based on gray-scale gradient, then considered the sharp of cross-border, we paired and markded the cross points. Finally, we removed the frame line and retained information of the intersection between frame line and character.3. Tilt correction of the numeral strings. Different person has different writing habit, many numeral strings are tilted. In order to facilitate the segmentation and make the features of character more focused, we need to normalize the direction. To calculate the tilt angle, we first rasied an algorithm based on the direction of digital stroke gradient, which is effectively for 90% numeral strings. For the other 10% cases, we rasied another algorithm based on the information of character frame. Two methods are complementary to each other, and achieved good results. 4. Cutting the string of numbers into single numeric characters. First, according to the contour information of each domain character contour, we determine the domain one number or touched numeral string. We rasied an algorithm based on gray-scale image to segmentate the touched numeral string. On the gray image, we look for the segmentation points based on the changes of gradient direction on the edge point of character, then, considered the information of different segmentation lines, we connected segmentation points and calculate the rate of reliability. And then we selected different segmentation lines to product different cutting combinations. Finally, we choose the best cutting combination due to the feedback of recognition.Combined with the k-nearest neighbor classifier, we achieved the recognition system of notes handwritten numeral strings. We tested 1000 digital images, each of which included 12 numeric characters, and the correct rate of recognition is 90.2%.