Complex-Valued Neural Network And Its Application In Traffic Sign Recognition

The parameters of traditional neural network are all real-valued, so it is hard for this kind of neural network to take care of both time and accuracy for the multistate recognition problem. Complex-valued neural can be multistate which makes it good at dealing with multistated patterns. For traffic sign recognition problem, neural network is one of most widely used methods. The traffic sign has two kinds of information: shape information and color information (red, yellow, blue, white, black). It is proper to convey these information with a gray-scaled image which can be seen as multistated patterns, so it is reasonable to address the recognition problem using complex-valued neural network. This thesis mainly focuses on the following three parts.Firstly, this thesis analyzes the basic structure of the complex-valued neural network. illustrates the disadvantages of the complex-valued S function, studies the complex-valued neural functions, and presents their basic characteristics and applications; it also compares the classification ability of real-valued perceptron with complex-valued one; it demonstrates the complex-valued BP algorithm and Hebbian learning rule.Secondly, this thesis designs the traffic-sign-oriented complex-valued neural network and its learning rules according to the characteristics of traffic signs. It employs the complex-valued step function to build neural model, then constructs network structure and decides the number of neurons; it adopts complex-valued Hebbian rules and inner product way to calculate the net weight matrix; then it analyzes the convergence of the network by taking advantage of the definition of energy function. In the image pre-processing stage. 2-dimensional Fourier Transform and Euler's formula are used, respectively, to transfer a gray-scaled image into phase information needed by the complex-valued neural network when stores the traffic sign. In the image post-processing stage, phase inverse transformation is used to change the phase image to gray-scaled image. Last, this thesis simulates the system. In order to realize recognition system with Labview, the Complex-valued step function module, phase-gray conversion module, traffic sign image reading module, and recognition algorithm module are programmed in Labview environment. The thesis also prensent the reason why "noise" appears in the recognition, compares and analyzes the performance of the real-valued network with the complex-valued network in traffic sign recognition problem, and demonstrates the advantage of the traffic sign recognition system using complex-valued neural network.