Aiming at the large-scale group decision problem LGDP a feature extraction method for decision makers?? evaluation information was proposed under the fuzzy environment of multi-granularity unbalanced hesitant language. Considering that the weights of different clusters formed by decision makers after clustering are affected by the differences of their decision makers a discrete coefficient is defined to correct the weights between clusters. Firstly the multigranularity language provided by decision makers was unified and the decision matrix with probability information was obtained. Secondly in computer vision analysis any image is composed of RGB three primary colors and the image is easier to extract feature than the data so the corresponding color image was obtained by calculating the RGB corresponding to all probability data in the decision matrix. The feature extraction algorithm was used to extract the characteristics of the evaluation information in the decision matrix which avoided the shortcomings of the existing decision methods that are difficult to extract the key features in the decision matrix quickly and effectively and were more efficient and concise in dealing with large-scale decision problems. After that the decision makers were further clustered to obtain different clusters and the revised weight between clusters was obtained by using the newly defined discrete coefficient. Then the final decision result was obtained by calculating the net flow size to sort the schemes. Finally the effectiveness and feasibility of the method were illustrated by an example of railway line scheme selection.