This paper points out the lack of autonomous behavior and independent reasoning ability of existing artificial intelligence models, and the limitations of data input methods that depend on explicit queries, and proposes a new machine thinking framework that mimics the way humans utilize mental images. This framework consists of three auxiliary units: an input data unit, a desire unit, and a mental image unit centered on a cognitive thinking unit, and uses natural language sentences or picture sketches as data to provide information and make decisions. In this paper, we present and discuss the results of verification experiments on the proposed framework.