Document Image Classification: Towards Assisting Visually Impaired

Our work aims to enable the visual information in a document to be accessible by the visually impaired or the blind people. The blind people prefer arts over science subjects in higher education because conveying equations and algorithms to them is seen as difficult. They should not be deprived of acquiring knowledge due to physical disabilities. In this work, we identify different types of images in an algorithm textbook as graphs, algorithm images, equations, and network flow diagrams with 98% accuracy. The Convolution Neural Networks used can predict new data instances with 99% accuracy except for recurrence tree images. We want to extend our work to extract the textual information in these images and produce image descriptions as alt text. This can produce a better speech output to assist the blind people.