混合神经网络应用于图象处理的研究
|
摘要
图象处理技术包括图象复原、图象压缩、图象分割、图象增强等一系列分支，目前它在遥感、字符识别、射线底片等众多领域得到迅速应用和推广。图象处理技术增强了人类对大千世界知识获取的能力，例如卫星云图的摄制、医疗监控系统的运作，乃至数字化视频光盘的压制，都随着图象处理技术的发展而逐步更新。
近几年，随着神经网络理论的深入研究，神经网络技术的并行计算能力、非线性映射和自适应能力等优点得到了国内外学者的充分重视，各种神经网络模型在图象处理领域中得到了广泛的应用。
本文将立足于神经网络技术和图象处理技术实际的结合情况，做出新的探讨和尝试。
论文的主要内容如下：
首先，作者系统的介绍了图象处理和图象编码的特点及其发展过程。全面、系统地概述了传统编码算法的优点，并同时指出了其存在的若干问题，为本论文指明了主要的研究方向。
然后，作者对各种流行的神经网络模型在图象处理领域中的应用进行了汇总，根据图象处理的具体内容对这些应用进行分类叙述，阐明了神经网络技术在图象处理领域中的优点和不足之处。
再次，作者依据自组织特征映射(SOFM)理论和主元分析(PCA)理论，提出了一种基于PCA／SOFM混合神经网络的图象压缩编码算法，并对SOFM网络学习参数的优化进行了探讨。实验证明，与普通PCA+SOFM连续编码算法和基本SOFM算法相比，这种混合编码算法由于占用存储空间少，因而降低了码书设计的计算量，并改善了码书的性能。
另外，作者针对上述基于神经网络的图象压缩软件方法，提出了硬件化结构方案，并探讨了其具体实现过程的要点和难点。
最后，作者对全文进行了概括性总结，并指出了其他有待深入研究的问题和今后工作的重点。
Image processing technology includes image restoration, image compression, image segmentation, image enhancement, etc. Currently, it is widely adopted in and spread to fields such as remote sensing, character recognition, radiographic film and so on. The technology enlarges the utilization of images, which we obtain from the boundless universe. And we can design digital video disc, satellite nephogram cemara or medical supervise system with high-performance based on the image technology.
Artificial neural networks technology research has gone a long way in the past decade. The specialists at home and abroad in the image processing field have paid high attention and been engaged in the advantages of Neural Network techniques such as the abilities of parallel computing, nonlinear mapping and self-adaptiveness, and applied a variety of Neural Network models into the image processing field.
According to the. practical combination of the neural networks technology and imge processing technology, this thesis researches into new algorithms in detail.
The main contents of this thesis are as follows:
Firstly, the author gives an overview on the characteristics and history of the image processing technology and image coding technology and introduces some basic coding algorithms in detail. The author also discusses on some problems on application, which will be partly settled in this thesis.
And then, the author reviews many applications of Neural Network in image processing and discusses the status quo and prospect of Neural Network. All these applications are categorized according to the phases of image processing and discussed in detail. In the conclusion of paper, he lists several disadvantages of Neural Network techniques.
Combining the Self-organizing feature map theory (SOFM) and Principle component analysis theory (PCA), the author proposes an image compressing algorithm based on PCA/SOFM hybrid neural network, which has the advantages of both PCA and SOFM. A new method of selecting initial codebook and distortion
criterion is presented to improve the efficiency of SOFM neural network according to the statistic feature of PCA transformational coefficient. Simulation results show that compared to successive PCA and SOFM algorithm or basic SOFM algorithm, PCA/SOFM hybrid algorithm has many advantages: less memory occupation; substantial reduction in computation and the better performance of codebook.
What is more, the author probes into hardware-based neural network, which he has engaged with image compression above.
Finally, the author makes a conclusion and proposes the future research directions.
近几年，随着神经网络理论的深入研究，神经网络技术的并行计算能力、非线性映射和自适应能力等优点得到了国内外学者的充分重视，各种神经网络模型在图象处理领域中得到了广泛的应用。
本文将立足于神经网络技术和图象处理技术实际的结合情况，做出新的探讨和尝试。
论文的主要内容如下：
首先，作者系统的介绍了图象处理和图象编码的特点及其发展过程。全面、系统地概述了传统编码算法的优点，并同时指出了其存在的若干问题，为本论文指明了主要的研究方向。
然后，作者对各种流行的神经网络模型在图象处理领域中的应用进行了汇总，根据图象处理的具体内容对这些应用进行分类叙述，阐明了神经网络技术在图象处理领域中的优点和不足之处。
再次，作者依据自组织特征映射(SOFM)理论和主元分析(PCA)理论，提出了一种基于PCA／SOFM混合神经网络的图象压缩编码算法，并对SOFM网络学习参数的优化进行了探讨。实验证明，与普通PCA+SOFM连续编码算法和基本SOFM算法相比，这种混合编码算法由于占用存储空间少，因而降低了码书设计的计算量，并改善了码书的性能。
另外，作者针对上述基于神经网络的图象压缩软件方法，提出了硬件化结构方案，并探讨了其具体实现过程的要点和难点。
最后，作者对全文进行了概括性总结，并指出了其他有待深入研究的问题和今后工作的重点。
Image processing technology includes image restoration, image compression, image segmentation, image enhancement, etc. Currently, it is widely adopted in and spread to fields such as remote sensing, character recognition, radiographic film and so on. The technology enlarges the utilization of images, which we obtain from the boundless universe. And we can design digital video disc, satellite nephogram cemara or medical supervise system with high-performance based on the image technology.
Artificial neural networks technology research has gone a long way in the past decade. The specialists at home and abroad in the image processing field have paid high attention and been engaged in the advantages of Neural Network techniques such as the abilities of parallel computing, nonlinear mapping and self-adaptiveness, and applied a variety of Neural Network models into the image processing field.
According to the. practical combination of the neural networks technology and imge processing technology, this thesis researches into new algorithms in detail.
The main contents of this thesis are as follows:
Firstly, the author gives an overview on the characteristics and history of the image processing technology and image coding technology and introduces some basic coding algorithms in detail. The author also discusses on some problems on application, which will be partly settled in this thesis.
And then, the author reviews many applications of Neural Network in image processing and discusses the status quo and prospect of Neural Network. All these applications are categorized according to the phases of image processing and discussed in detail. In the conclusion of paper, he lists several disadvantages of Neural Network techniques.
Combining the Self-organizing feature map theory (SOFM) and Principle component analysis theory (PCA), the author proposes an image compressing algorithm based on PCA/SOFM hybrid neural network, which has the advantages of both PCA and SOFM. A new method of selecting initial codebook and distortion
criterion is presented to improve the efficiency of SOFM neural network according to the statistic feature of PCA transformational coefficient. Simulation results show that compared to successive PCA and SOFM algorithm or basic SOFM algorithm, PCA/SOFM hybrid algorithm has many advantages: less memory occupation; substantial reduction in computation and the better performance of codebook.
What is more, the author probes into hardware-based neural network, which he has engaged with image compression above.
Finally, the author makes a conclusion and proposes the future research directions.
引文
[1] G. A. Baxes, "Digital Image Processing: Principles and Applications", Wiley, New York, 1994.
[2] IEEE Std 610. 4-1990, "IEEE Standard Glossary of Image Processing and Pattern Recognition Terminology", IEEE, New York 1990.
[2] IEEE Std 610. 4-1990, "IEEE Standard Glossary of Image Processing and Pattern Recognition Terminology", IEEE, New York 1990.