Improving the accuracy of image caption extraction is one of the problems related to computer vision and depends on natural language processing. Therefore, the image captioning problem is performed by many different combination methods. This paper proposes a method to build captions for multi-object images by combining attention mechanisms to focus on the main objects and relationships in the image; thereby building a knowledge graph and combining semantics from ConcepNet. To do this, each object in the input image is recognized and classified using a deep learning network. The objects and relationships between the objects of interest will be added to the knowledge graph. Finally, semantics from the built knowledge graph and ConcepNet are combined to complete the caption for the input image. The method is tested on MS-COCO and Flickr30k image datasets with BLUE-1 and BLUE-4 metrics to evaluate these results. The experimental results are compared with the results of other methods on the same dataset; this proves the feasibility, correctness, and efficiency of the implemented method.