Pavement Distress Perception using Fused Convolutional Neural Networks

Abstract

The issue of highway maintenance is becoming increasingly prominent. Achieving high-precision detection of cracks and various types of pavement distress through technical means is fundamental to intelligent road maintenance. Firstly, this paper studies convolutional neural networks (CNNs) for pavement distress detection in highway scenarios. The three neural network models Faster-RCNN, YOLOv8s, and DeepLabv3-MobilenetV3 are comparatively evaluated for their effectiveness in these tasks. Secondly, a new network architecture, named Fused Convolutional Neural Networks for Pavement Distress Perception (FCNN-PDP), is proposed. In this architecture, the C2F-Damage module from the FasterNet Block replaces the original C2F module in the model neck, and a BiFPN-Neck structure is adopted as the neck network of FCNN-PDP. Finally, comparative experiments of FCNN-PDP have been carried out, and the experimental results show better performance of the proposed FCNN-PDP network in detection of pavement distress in highway scenarios. The test accuracy and time consumption of FCNN-PDP are 98. 88% and 20.20 ms, respectively, and the ablation experiments show the effectiveness of improving modules in FCNN-PDP.