A study of high-level recognition integrated system realizing scene understanding by using combination of coarsely- segmented image regions

我们已经开发了一种图像处理模型和算法，其中分割和提取粗区域而不考虑精细结构，并且基于分割区域之间的关系来识别和理解包括多个对象的场景图像。我们已经开发了实现这些算法和集成系统的实时高阶图像识别的超大规模集成电路。在所开发的系统中，粗区域分割的连续融合网络和弹性图匹配（EGM）在FPGA与PC通过USB连接。在粗分割区域中，当一个目标被另一个目标部分隐藏并分成两个区域时，我们成功地利用Terman-Wang非线性振荡器网络的同步触发，通过考虑区域的亮度，同时提取出两个分离的区域。在此基础上，提出了一种基于简化振荡器网络模型的细胞自动机型区域提取算法，并设计了实现该算法的像素并行数字电路，在FPGA上实现了该算法。我们已经提出了两种数字和模拟类型的像素，提出了一种用于生成主观轮廓的并行各向异性扩散算法和电路，这些算法已分别在CMOS数字电路和合并式A/D电路中实现。使用这些扩散算法，我们已经成功地产生了主观轮廓的典型人物，如Kanizsa的三角形。

We have developed an image processing model and algorithms, where coarse regions are segmented and extracted irrespective of fine structures, and a scene image including plural objects is recognized and understood based on the relationship between segmented regions. We have developed VLSIs implementing these algorithms and an integrated system for real-time higher-order image recognition. In the developed system, coarse region segmentation by a resistive-fuse network and elastic graph matching (EGM) are performed in FPGAs connected with a PC via USB. Related fundamental technologies have also been developed.In coarsely segmented regions, when an object is partly hidden by another object and separated into two regions, we have succeeded to extract two separated regions simultaneously by considering the brightness of the regions when synchronous firing in Terman-Wang nonlinear oscillator networks is used. Furthermore, we have proposed a cellular-automaton-type region extraction algorithm based on a simplified oscillator network model, and a pixel-parallel digital circuit performing this algorithm, which has been implemented in an FPGA.In order to compensate for lacking of edge information in coarse region segmentation, we have proposed two digital and analog-type pixel-parallel anisotropic diffusion algorithms and circuits for generating subjective contours.These proposed algorithms have been implemented in a CMOS digital circuit and merged A/D circuit, respectively. Using these diffusion algorithms, we have successfully generated subjective contours for typical figures such as a Kanizsa's triangle.