A CNN-Based Length-Aware Cascade Road Damage Detection Approach

Area of Science:

• Computer Science
• Artificial Intelligence
• Civil Engineering

Background:

• Road damage detection is crucial for public transportation safety.
• Current deep convolutional neural networks (CNNs) struggle with weak semantic information and abnormal geometric properties of road damage.
• Suboptimal feature representation leads to inadequate detection results.

Purpose of the Study:

• To propose a novel CNN-based cascaded damage detection network, CrdNet.
• To address the limitations of existing methods in representing weak semantic information and abnormal geometric properties of road damage.
• To enhance the accuracy and robustness of road damage detection.

Main Methods:

• Introduced LrNet, a novel backbone network for high-to-low level feature fusion, improving weak semantic information representation.
• Applied multi-scale and multiple aspect ratios anchor mechanisms to generate high-quality positive samples for abnormal geometric properties.
• Designed an adaptive proposal assignment strategy with cascade predictions for varied range dependencies.

Main Results:

• The proposed CrdNet achieved a mean average precision (mAP) of 90.92% on a collected road damage dataset.
• Demonstrated superior performance in localizing and classifying road damage compared to existing methods.
• Validated the model's robustness in handling diverse road damage characteristics.

Conclusions:

• CrdNet effectively addresses the challenges of weak semantic information and abnormal geometric properties in road damage detection.
• The novel network architecture and strategies significantly improve detection performance and robustness.
• The findings contribute to advancing road safety through more accurate automated damage assessment.