The region proposal search is one of the most active research topics of machine vision. The low efficiency of object detection using the traditional exhaustive search can be improved by optimizing the accuracy of search algorithm. The Girvan-Newman (GN) splitting for community discovery in complex networks is introduced as well as the features of small object regions. A novel method to generate small object regions is proposed by using the network structure of image. The algorithm constructs the relationship between the images and the graphs based on the similarity of color histograms between regions. It can obtain possible regions through the generation of connected subgraphs. This algorithm can meet the higher recall rate in the case of generating fewer candidate regions and further optimize the time consumption of small object detection.