A Software for Rapid Annotation of Scene Objects Based on Saliency Object Ranking

With the development of artificial intelligence (AI), a mainstream research line in building a non-invasive intelligent system for assisting people with visual disabilities is to capture images with a camera, and then using AI to identify the content of images and convert visual information into auditory information. The key point is to establish high-quality personalized datasets based on real open scenes for AI model training. However, most of existing annotation software is designed for face recognition, automatic driving and other tasks, whose annotation information type is monotonous. Also, when converting an image with multiple objects into audio information, the order of objects needs to be determined. Few studies have focused on this issue currently. Therefore, we develop an image annotation software for rapid construction of small sample databases of AI models that can be used in intelligent blindness assistance systems. The software can additionally mark information such as size and affiliation. Moreover, it has a database module to facilitate the management of multiple annotation tasks. Then, a saliency object ranking network based on multi-task cascade is designed. The object ranking results can provide a reference for the sequence of audio information. The Hit Rate, Relative Ranking accuracy and Saliency Object Ranking score of the ranking network are better than that existing saliency object network on the validation set. Finally, the network is deployed into the software to realize the AI-assisted annotation, and to quickly establish the personalized dataset.