DeepUnseen: Unpredicted Event Recognition Through Integrated Vision-Language Models

Deep Learning-based segmentation models provide many benefits for scene understanding. However, such models have not been used and tested for unpredicted events like natural disasters by hurricanes, tornados, and typhoons. Since low illumination, heavy rainfall, and storms can degrade image quality, implementing a single state-of-the-art (SOTA) model only may fail to recognize objects correctly. Also, there are more enhancements to segmentation that remain unsolved. Thus, this paper proposes a vision-language-based DL model, namely, DeepUnseen, by integrating different Deep Learning models with the benefits of class and segmentation. Experimental results using disaster and traffic accident scenes showed superiority over a single SOTA Deep Learning model. Moreover, better semantically refined classes are obtained.