For image captioning such as for computer game images or other images, bottom-up attention is combined with top-down attention to provide a multi-level residual attention-based image captioning model. A residual attention mechanism is first applied in the Faster R-CNN network to learn better feature representations for each region by taking spatial information into consideration. In the image captioning network, taking the extracted regional features as input, a second residual attention network is implemented to fuse the regional features attentionally for subsequent caption generation.