Multi-class Concrete Defect Detection Algorithm Based on Rep-PVT

Concrete defects are usually overlapped by multiple defects, making accurate identifying of multi-class defects still a challenging task. In this study, we propose a multi-class concrete defects detection algorithm named Rep-PVT, which is based on residual Pyramid Vision Transformer (PVT). The Rep-PVT consists of two channels: one channel utilizes the RepVGG network for feature extraction, while the other channel employs the PVT. The features of the two channels are then merged, and the results are output through a Fully Convolutional Networks layer, enabling multi-class defects detection on concrete surface. Compared with MetaQNN [8], VGG [10], Texture-CNN [11], WRN [12] and DenseNet [13], the proposed Rep-PVT has achieved better precision in both the validation set and the test set on the CODEBRIM dataset. Specifically, compared to the PVT, our method exhibited an improvement of 3.98% and 1.43% in precision on the validation and test sets respectively.