Skip to the main content

Original scientific paper

https://doi.org/10.31803/tg-20231018184747

Improving Freedom of Visually Impaired Individuals with Innovative EfficientNet and Unified Spatial-Channel Attention: A Deep Learning-Based Road Surface Detection System

Amit Chaudhary ; Harcourt Butler Technical University, Nawabganj, Kanpur, Uttar Pradesh 208002, India *
Prabhat Verma orcid id orcid.org/0000-0001-9193-7744 ; Harcourt Butler Technical University, Nawabganj, Kanpur, Uttar Pradesh 208002, India

* Corresponding author.


Full text: english pdf 1.469 Kb

page 17-25

downloads: 403

cite


Abstract

Individuals with visual impairments often encounter substantial challenges navigating outdoor spaces due to their inability to perceive road-surface conditions. This study introduces an innovative method that harnesses deep learning to identify and categorize road surfaces, aiming to enhance the independence and mobility of the visually impaired. Leveraging the EfficientNetB0 model as a foundational framework and employing unified spatial-channel attention, we classified road surface images captured from a wearable camera. Through rigorous training and evaluation on a substantial dataset of road images, our modified system exhibited remarkable performance, accurately identifying road surfaces with an impressive 99.39% accuracy rate. This deep learning-driven approach holds promise as a pivotal tool for improving the autonomy and safety of individuals with visual challenges by providing instantaneous feedback on road conditions.

Keywords

attention mechanism; deep learning network; EfficientNet-B0; pedestrian with vision limitations

Hrčak ID:

327593

URI

https://hrcak.srce.hr/327593

Publication date:

14.3.2025.

Visits: 892 *