Automatic image captioning has many important applications, such as the depiction of visual contents for visually impaired people or the indexing of images on the internet. Recently, deep learning-based image captioning models have been researched extensively. For caption generation, they learn the relation between image features and words included in the captions. However, image features might not be relevant for certain words such as verbs. Therefore, our earlier reported method included the use of motion features along with image features for generating captions including verbs. However, all the motion features were used. Since not all motion features contributed positively to the captioning process, unnecessary motion features decreased the captioning accuracy. As described herein, we use experiments with motion features for thorough analysis of the reasons for the decline in accuracy. We propose a novel, end-to-end trainable method for image caption generation that alleviates the decreased accuracy of caption generation. Our proposed model was evaluated using three datasets: MSR-VTT2016-Image, MSCOCO, and several copyright-free images. Results demonstrate that our proposed method improves caption generation performance.
Keywords: deep learning; image captioning; motion estimation; object detection.