ON A MODELING AND APPLICATION OF HIGH LEVEL VISUAL SEARCH

近年来，许多研究开始使用视觉系统来辅助用户或中央系统的高层推理和规划。视觉搜索在这些应用中是一个非常重要和基本的功能，本研究构建了一个视觉搜索神经网络模型，使视觉搜索能够更高层次和更灵活地到达中央系统，扩大了视觉搜索的应用范围。该模型输出目标物体的形状、颜色和运动方向等高层概念。该模型不同于传统的神经网络中信息从输入层传播到输出层的双向传播。根据该机制，该模型在加入目标的部分概念时，抑制不必要的背景信息，并选择性地处理必要的信息。然后，该模型不仅能够稳定地识别目标，而且使用高级概念提供了灵活的接口。此外，在这方面，

Recently, many researches, which use a vision system to aid high level reasoning and planing of users or central systems, are started. The Visual search is a very important and basic function in those applications.In this research, we constructed a visual search neural network model, which enable a high level and flexible visual search to the central system, and enlarged the applications of the visual search. The proposed model outputs high-level concepts such as shape, color and motion's direction of target objects. This model differs from conventional neural networks, in which information is propagated from input-layr to output-layr, because of bidirectional information propagation. According the mechanism, the model inhibits the unnecessary background information and processes necessary information selectively, when partial concepts of the target are put into it. Then, the model not only recognizes objects stably, but also provides flexible interface using high level concepts. Furthermore, in this