基于神经网络的信元优化调度和相位特征识别

【作者】 李现国；

【导师】 申金媛；

【作者基本信息】 郑州大学 ， 物理电子学， 2006， 硕士

【摘要】 人工神经网络在优化控制和模式识别方面得到了广泛的应用。本文对人工神经网络方法在ATM(Asynchronous Transfer Mode，异步传输模式)信元优化调度和三维物体相位特征识别中的应用进行了研究。 1．基于神经网络的信元优化调度：在多重队列中采用每条入线在同一个时隙内可传送多于一个信元的策略，提出了两种新的Hopfield神经网络的能量函数，进而利用Hopfield神经网络实现对信元的控制和调度。计算机仿真模拟比较表明，我们所提出的这两种方法均提高了吞吐率，减少了信元时延，并降低了信元丢失率，消除了队头阻塞造成的性能恶化，提高了ATM交换结构的性能，实现了信元整体上的优化排程控制。由于Hopfield神经网络易于电子或光电技术实现，交换结构和缓冲器也无需加速，因此，这两种方法不失为有效的信元优化调度方案。 2．基于神经网络的相位特征识别：提出一种基于人工神经网络的三维物体相位特征识别方法。首先利用波长扫描数字全息技术提取物体的相位特征，然后将物体的这些相位特征作为学习模式训练一个BP神经网络，最后利用训练好的网络对三维物体进行识别，计算机模拟表明，对于具有小尺度变化的透明、半透明三维物体识别，该方法的正确识别率为100％，而且还具有一定的容错性。这为透明或半透明三维物体进行小尺度不变性识别提供了一个新的方法。更多还原

【Abstract】 Artificial neural network has been used for solving optimal control and pattern recognition problems extensively. In this dissertation, we concentrate our attention to applying artificial neural network in both cell optimal scheduling in ATM environments and phase feature recognition of 3-D objects.1. Cell optimal scheduling based on neural network: Two new energy functions of Hopfieid neural network are employed based on the multiple input-queuing cooperating with the policy of more than one cell transferred in each input line during every time slot. And then we use the Hopfieid neural network to control and schedule the cells. The computer simulation results show that our approaches not only greatly improve the throughput but also lower down the cell loss probability and reduce the average latency of ATM switching fabrics. That is the performances of ATM switching fabrics are improved due to reducing the head of line blocking (HOL blocking).Considering parallel information processing, easy accomplishment by electronic or optoelectronic technique of Hopfieid neural network and the ATM Switching Fabric and the buffer without speedup, the two approaches proposed in this dissertation are effective cell scheduling schemes.2. Phase feature recognition based on neural network: A new approach based on phase features combined with neural network model is proposed for recognizing 3-D objects in this dissertation. First, the phase features of a transparent or semitransparent 3-D object were extracted by wavelength-scanning digital holography and numerical reconstruction technique. Then, A BP neural network trained by those reconstructed images including the phase features of 3-D objects was used to recognize the input objects. The computer simulations show that the correct recognition rate is up to 100% for the training objects or ones with some scale variance. It demonstrates that the method we proposed is effective. And it gives a new way for recognizing the transparent or semitransparent 3-D objects with some scale invariance.更多还原