www.bewet.fora.pl

loczytzn

Thermal explosion of gunpowder standard materials of the standard value of uncertainty

Layer, the output target decision. Due to the recognition of the characters are 0 to 9 and A ~ D a total of 14 characters, authors used a 8421 yards to coding. Output \the number of nodes is 4. Basic BP neural network learning for a long time, the problem of slow convergence rate, of the following improvement of BP algorithm. First, the use of approved training to update the weights, that is, a sample of each load, calculate the weight change, but one does not immediately update the weights, until all training samples are loaded once before the change in the cumulative weight, Update weights. Second, increase the momentum term and dynamic adjustment of learning rate. Increase the momentum of items such as formula (3) below: AW (t) a q3X + aAW (t-1) (3) where: w on behalf of a layer weight matrix; X on behalf of a layer of input vector; OL for the momentum factor, generally OL ∈ (O, 1). Momentum reflects the adjustment of previously accumulated experience of the adjustment for time t from the damping effect. When the error occurred suddenly ups and downs, it can reduce the oscillation tendency to improve the training speed. Learning step can be changed by the following method: first to set an initial learning rate,[link widoczny dla zalogowanych], if the weight adjustment after one batch after the total error increases, then this adjustment is invalid, and a flv (fl 0). 3.2 Design of the attention points as the transfer function of BP neural network is Sigmoid function, this function will limit the output values (O, 1) range, in fact, the output value can never be fully equal to 0 or 1. Close to 0 and 1 because the two regions, functions relatively flat, and its derivative tends to 0, each learning time,[link widoczny dla zalogowanych], the value of the correction will be minimal, so spent most of the time. Thus, in learning as much as possible without affecting the accuracy of the premise, the expected value of the output becomes 0 to 0.1,1.0 0.9, so the standard output vector becomes a combination of 0.1 and 0.9, out of the function of the flat zone, significantly faster network convergence, derivative into the possibility of the region close to 0, thereby reducing the number of iterations, saving time. In addition, when a network after a period of training, I always find most of the output is equal to or close to the ideal value of the network, leaving a large input output error only a small part of the input samples. In other words, after learning later, the network already has a certain amount of recognition, can identify most of the characters, only part of the network can not correctly identify the characters. Taking into account a feature of BP network, the new input of the sample will change the weights of the network,[link widoczny dla zalogowanych], affecting to the previous sample, so in actual operation, the output error can not be extracted a large sample of individual learning, but in the previous sample increase in the output error in a large number of study samples, this will improve the recognition rate. 4 The results of the VC + +6.0 using the recognition software was developed by a large number of images collected from the training and testing, in addition to a number of shooting does not recognize the blurred image, other image recognition results are more satisfactory. Rejection of those characters, we can use interactive methods of characters entered the correct result. Identification of software interface running as shown in Figure 2. JJj. . . JJ called. Figure 2 recognition software running interface [1] Zhang Sri Lanka. Chinese character recognition technology to EM]. Beijing: Tsinghua University Press,[link widoczny dla zalogowanych], 1992. E23 Han Liqun. Artificial neural network theory, design and application [M]. Beijing; Chemical Industry Press,[link widoczny dla zalogowanych], 2002. [3] Hu Xiaofeng, Zhao Hui. Visualc + + / MATLAB image processing and recognition of the practical case selection of EM]. Beijing: People's Posts and Telecommunications Press, 2004. [4] Ma Shiping. Bank money box implementation of character recognition system [J]. Computer Engineering. 2003, (11): 173 ~ 174. [5] Li Xiaoping. BP algorithm using the container number identification [J]. Beijing University of Technology. 2001, (6): 345 ~ 347. I-6] RichardG. Casey, EricLecolinet. ASurveyofMethodsandStrategiesinCharacterSegmentation [J]. IEEETRANSACTIONSONPATERNANALYSISANDMACHINAEINTELLIGENCE. 1996.690 ~ 706.

相关的主题文章：

[link widoczny dla zalogowanych]

[link widoczny dla zalogowanych]

[link widoczny dla zalogowanych]