In this paper, a mechanism for large scale image/video retrieval on the distributed environment is proposed, and an approach of image/videopsilas automatic annotation considering semanticspsila tolerance relation between pre-defined classes is described. The image/videospsila semantics is represented by a vector of associative values with the pre-defined classes which is generated by the distributed nodes providing the contents of image/video. On the other hand, by using bidirectional associative memories, the method how to generating associative values with pre-defined classes regarding different dimensions is also depicted. Image/videospsila retrieval results based on the pre-defined classes shows the effectiveness of the proposed methods.