Evaluations on Speech/Image Coding and Recognition Systems from Information Theoretic Point of View

存在各种各样的信息源：电话号码的数字序列是无冗余源的一个例子，而传统的语音信号对应于具有大量冗余的一个数量级的源。在本项目中，我们首先尝试通过设计一个通用编码器对这些大范围冗余的源进行编码，然后从信息论的角度对该编码和识别系统进行评估。得到以下结果：系统的整体失真我们关注一个有源有输出的系统，然后我们定义了一个涵盖以下项目的整体失真；源信号与被检索的模拟/数字信号的互信息I（X,Ya）和I(X,Yd)，量化误差，识别误差2。为了提高分形冗余源的识别率，设计了一种广义非对称纠错码。利用该非对称纠错码，对识别率的性能进行了评价。当将手写数字字符作为图像数据处理时，它们具有大量的冗余，但是作为特定的数字序列，它们没有冗余。数字字符识别系统可以被建模为一个有限错误模式的通信信道。针对该信道设计了纠错码，提高了数字识别率。

There exist various kinds of information sources: Numeric sequences of telephone numbers are an example of no redundant sources, while conventional voice signals correspond to the source with a plenty of redundancy of the order of one magnitude. In this project, we first attempt to code these sources of wide range of redundancy by designing a universal coder, and then evaluate this coding and recognition system from information theoretic point of view. The following results are obtained;1. Overall distortion of system We focused on a system with a source and an output, and then we defined an overall distortion covering the following items; mutual information I(X,Ya) and I(X,Yd) of the source signal and the retrieved analog/digital signal, quantization error, recognition error.2. Improvement of recognition rate based on asymmetric error correcting codes For improving recognition rates with a fractal redundancy source, we designed a generalized asymmetric error correcting code. By using this asymmetric error correcting code, the performance of recognition rate was evaluated.3. Error correcting code for limited error pattem Hand-written numeric characters have a plenty of redundancy when one treats them as image data, however, they have no redundancy as a specific numeric sequence. The numeric character recognition system can be modeled as a limited error pattern communication channel. We designed an error correction code for this specific channel, and the recognition rate of numbers is concluded to be improved.