Development of a high-speed image-understanding system designed directly from image data

本项目的目标是实现一种在常规技术下无法实现的高速图像理解系统。为了实现该系统，我们实现了直接从图像到VLSI的电路设计。并将这种设计方法应用于“帕尔岑窗方法”和“概率神经网络方法”。在这两种方法中，都使用了使用大量样本图像构造的函数，并且直接根据图像样本数据来设计函数。由于这种数据直接实现，函数的计算速度比传统方法要快得多。利用可重构LSI(FPGA)建立了原型系统，并获得了以下结果：1)利用遗传算法对Parzen窗方法中的窗函数进行了优化。2)数据直接实现的思想不仅有效地应用于图像识别，而且也有效地应用于声纳频谱识别。3)开发了基于“概率神经网络方法”的图像理解系统，并与PC机连接。使用PC机对系统进行了有效和准确的评价，并利用PC机对系统的输出进行了主成分分析，以提高识别的准确性

The goal of this project is to realize a high-speed image-understanding system that has not been realized under the conventional technologies. In order to achieve the system, we implement the circuits designed directly from the images onto the VLSIs. And we apply this design approach to the "Parzen Window Method" and the "Probabilistic Neural Network Method". In both methods, the functions constructed using a great number of sample images are used, and the functions are directly designed from the image sample data. Because of this data-direct-implementation, the functions are calculated much faster than those used in the conventional approachesThe prototype system was established using reconfigurable LSIs (FPGAs) and following were obtained1)Window functions in the Parzen Window Method were optimized using the GA (genetic algorithm). High recognition accuracy was obtained with the window functions in the face recognition problem2)The proposed idea of the data-direct-implementation was efficiently applied not only to the image recognition but to the sonar-spectrum recognition3)The image-understanding system based on the "Probabilistic Neural Network Method" was developed and is connected with a PC. 'The system was effectively and precisely evaluated using the PC. The PC was also used to apply the principal component analysis to the outputs from the system in order to improve the recognition accuracy