Recognition For Handwritten Numbers Of Bank Notes Based On Neural Network Optimized By Genetic Algorithm

At present, the technology of handwritten number recognition is widelyapplied in many fields, where auto recognition of bank notes by computers is animportant topic, naturally recognition of amount and ID card numbers is a keypart. In order to solve this major problem, author presents the recognitionalgorithm of handwritten numbers on the bank notes. Detailed image processingwas introduced in the fifth charcper, including binarization, skew detectionand correction, method of mean filter, method of thinning, the segmentingmethod of connective digits, the DDA algorithm for separation of strokesand so on. The novel and interesting content of image processing usd in thesection 5 was simply introduced in this abstract. The sixth charcper introducesrecognition for handwritten numbers of bank notes based on BP neural networkoptimized by genetic algorithm. By contrasted with the different network, theexperiment has proved that this recognition algorithm is sufficient effectiveto recognition for handwritten numbers of bank notes.Two-dimensional weighted maximum entropy Author presents "two-dimensional weighted maximum entropy"on thefoundation of reference document, having considering special situationwhen background area differs greatly from aim area. By combining itwith improved genetic algorithm, we obtain the optimum two-dimensionalthreshold. Improvement of the two dimensional maximum entropy proposedby the reference document, can be summarized to two aspects by us :the improvement of the differentiation district and the improvemeddifferentiation function, where the latter can be defined in this paperby F ( s,t)= H(A)PA +H(B)PB.Naturally, PA and PB are probability ofthe background area and aim area respectively,while H (A)and H (B)aretwo-dimensional entropy of the two areas respectively.An improvemed method of skew correctionScaning may cause notes' skew in some degree. Based on the fact thatthe frame of character area is two pairs of parallel lines, we can easilydetect the parallel lines on the boundary by utilizing improvemed houghtransform, calculate the varied coordinate vale of each row and correct theskew at last. The experiment has proved much computing amount can besaved by using this algorithm compared with traditional algorithm.An improved method of median filterUsually,there is guass noise and isolated noise in many nature imagessimultaneously. It is difficult to getting rid of guass noise and isolated noise onlyby median filter or mean filter at the same time. To overcome the problem, theimproved method of median filter was adopted in this paper. This method appliesauto adapted operators on the N × N area of every point in the processedimage. The more the gray value be close to the median,the more the weightof operator is strong.DDA algorithm for separation of strokesSeparation of strokes is a general problm in the processed images afterimage processing above-mentioned. The DDA algorithm for generating lineswas used for solving the problem. When the smallest distance of strokes' end issmaller than the threshold, we connect the corresponding strokes.Use genetic algorithm to obtain optimum original weight ofBP neural networkThe choice of original weight in BP neural network usually effects theproperties of neural network and recognition result greatl. genetic algorithmwas used to obtain optimum original weight of BP neural network, where thefitness funtion is f = e+1 ε, and e is error square sum of all sample, εis a sufficient positive number.The method of weighted multi-networks combinationBecause of the complex degree of handwritten numbers, if we only useone kind of BP neural network to carry on the classification, the network will beextremely complex, and it would inevitably be able to affect the networkconvergence rate, the training effect and recognition rate. Thus this article usesfour BP network to carry on the work, each network only selects one kind ofcharacteristic as the input, here author uses the method of weighting tocombine various network value to take the recognition result, and each weightis established according to the recognition rate for same sample. Thus wecan not only simplify the complex degree of BP network, but also considerhandwritten numbers' typical characteristic overall, naturally the method canenhance the effect of recognition.Gray characteristic networkTwo-dimensional dispersed discrete cosine transform(DCT) was used inthis article before training the BP neural network to getting rid of image gray'sredundancy often generated by the above binarization. Because of discretecosine transform's characteristic, the shortcoming of slow convergence rate of BPnetwork resulted by redundancy will be overcomed greatly.Geometric structure networkThe geometric structure is handwritten numbers' stable and importanttrait, representating the characteristics of processed handwritten numbersat high level. Because of the property, we can take the number,diretionand position of the end, protruding points, three cross points and fourcross points as the input.Black pixels'distance networkIt is well known that black pixels' distance of each row is handwrittennumbers' stable trait, we can take distance of the last pixel and the firstpixel in the fixed row as the input value in this network, and the step wasoperated towards all row in the processed area.Scanning beams networkBecause of the handwritten numbers' traits, algorithm of scanning beamswas usually applied to avoid the bad effect may resulted by the characters'deformation. The operation is that we scan the image from the left border ,anddo it from left to right. At the same time we calculate the intersection points'number of the character and every scanning beam.