This thesis deals with one of the problem in high level vision. The very factor for reliable perception machine is the methodology for association of various sources of information rather than the measurement of information itself. In limited domain and on the limited measurements, some schemes for reliable vision system are presented. In uncertain image information the linear approximation of boundary shape representation is extracted and structure analyzer finds the inter-relations between subparts and describes them by di-graph. Upon these local feature the paradigm for reliable process, hypothesis and verification, is implemented. The global interpretation technique and utilization of contextual information will be discussed. Some error corrective procedure in verification phase is realized with feedback control strategy. The two components of environments for hypothesis and verification is working elements and record of state. The working elements, set of hypotheses, are the atomic item for reasoning and changes nonmonotonically. At each stage of matching, the resultant change of current state is changed. As a result, hypothesis and verification is an efficient approach for overcoming the wall of ambiguities in numerical taxonomy.