Lightweight deep learning for real-time road distress detection on mobile devices

Area of Science:

• Computer Vision
• Artificial Intelligence
• Infrastructure Engineering

Background:

• Manual road inspections are inefficient and time-consuming.
• Current automated systems require computationally intensive hardware, limiting scalability.
• Accurate road distress detection is vital for infrastructure maintenance and transportation safety.

Purpose of the Study:

• Introduce MobiLiteNet, a lightweight deep learning model for efficient road distress detection.
• Enable real-time, accurate road distress detection on mobile devices like smartphones.
• Improve road infrastructure management and intelligent transportation systems.

Main Methods:

• Developed MobiLiteNet, a lightweight deep learning approach for mobile deployment.
• Incorporated Efficient Channel Attention for performance enhancement.
• Applied structural refinement, sparse knowledge distillation, structured pruning, and quantization for computational efficiency.

Main Results:

• MobiLiteNet demonstrated improved performance over the baseline MobileNet model on mobile devices.
• The model achieved significantly reduced computational costs while maintaining high detection accuracy.
• Validated effectiveness using a diverse dataset from Europe and Asia.

Conclusions:

• MobiLiteNet offers a scalable, real-time solution for road distress detection on mobile platforms.
• The approach facilitates more efficient road infrastructure management.
• Contributes to advancements in intelligent transportation systems through accessible technology.