This paper addresses the challenge of creating a new system to estimate the impression of an image. The proposed system combines the human annotated tags of images and an image classification method to discover “showing a photo, what are people looking at?”. Then, to tackle the challenge “what are they thinking about the one they look at?”, the semantic association strengths between adjectives and image keywords are computed by pointwise mutual information (PMI) and the pattern frequencies using a machine learning approach. To select the output, we use a rank aggregation method, Borda's method, to generate an acceptable ranking for a given set of rankings and the top na adjectives (in this paper na is 5) are chosen according to the estimated values. The main contribution of this method is to design an effective method for estimating the association of the impression adjectives with images. We evaluated the proposed approach using two tasks: the first one is the performance of the task of keyword extraction and the second one is the efficiency of the proposed method. For the performance of the proposed method, we carried out subjective experiments and obtained fairly good results.