Multi-perspective feature collaborative perception learning network for non-destructive detection of pavement defects

Abstract

Pavement defects detection has made significant progress with the development of convolutional neural networks. Due to the topological complexity of pavement defects regions, most existing detection methods are limited to a limited number of types of defects, which results in unsatisfactory detection results for multi-classified defects. Several studies have used enhanced data to address this issue, however, few have considered the inefficiency caused by imbalanced differences between classes from the perspective of the network model itself. In light of this issue, our goal is to fill this gap by better exploring the fusion of different feature information to improve the performance of the network. As a result, we propose a multi-perspective feature collaborative perception learning network (MFCPLNet), which uses the model's contextual information to guide high-quality multi-classified defects detection and that can be adapted to various pavement conditions. Specifically, for adaptive feature extraction, an irregular perceptual extraction network (IPENet) with enhanced edge information is first designed to draw upon hybrid deformable convolution for efficient network block hybridization. Additionally, a multi-balanced collaborative learning mechanism (MCLM) consisting of auxiliary axial information and 3D feature guidance is constructed to enhance deep semantic features, improve the localization of defects in defective regions, and compensate for weak category supervision. Lastly, in the detection head, an iterative soft aggregation module has been designed to enhance scalable spatial interactions. The results of extensive testing were based on publicly available pavement defects datasets. Based on the proposed network, the mAP index reaches 29.6% and 33.8%, respectively, an increase of 8.3% and 4.6% over the baseline, which is qualitatively and quantitatively superior to mainstream methods.