Explainability to Image Captioning Models: An Improved Post-hoc Approach through Grad-CAM

Providing appropriate explanations to deep learning-based image captioning models is essential to enable users to trust the generated captions. In this paper, we present a novel approach for explaining the captions generated by deep image captioning models. Specifically, we propose a segmentation based approach utilizing a pretrained, state-of-the-art captioning model known as BLIP. Initially, queries are generated from the captions, followed by an image segmentation module based on RCNN, trained on COCO dataset. In the next step, a similarity score between the query and the segmented regions of the image is computed using the CLIP algorithm. In the final stage, an explanation module using the Grad-CAM approach tailored to the final attention layers of the CLIP model is incorporated, which generates trustworthy visualizations for the captions. Thus, apart from providing accurate segmentation maps of the queries, the proposed method provides additional saliency maps for better visual explanations. Experiments conducted on a diverse set of images from the COCO dataset demonstrate the effectiveness of our approach, yielding superior results in comparison to similar segmentation based caption explanation models.