A Vision for All: A Dataset and CNN Framework for Navigating Educational Spaces for the Visually Impaired

Over the years, numerous datasets featuring both images and text have been introduced, driving the development of innovative methods that integrate natural language processing and computer vision. Nonetheless, there remains a demand for datasets that present images within their authentic context. This paper presents a large-scale image dataset collected from Zayed University and the British University in Dubai, designed to support the development of AI-powered assistive technologies for visually impaired individuals in academic settings. The dataset includes 300,000 images representing university facilities such as classrooms, labs, and safety features. Images were captured using direct photography and video frame extraction, incorporating a range of conditions. Preprocessing techniques ensured the dataset's high quality and variability. CNN and ResNet50 classification models were applied, and they achieved 90% and 93% accuracy, respectively. This dataset paves the way for an envisioned intelligent Arabic voice-based navigation system, supporting both standard and Gulf Dialect Arabic. This paper details the creation process of the dataset, the dataset structure, and the potential contributions to advancing research in assistive technologies for individuals with visual impairments.