Pixel-level pavement crack detection using enhanced high-resolution semantic network

Abstract

Pixel-level crack detection is crucial in pavement performance assessment. Current deep learning-based detection methods first encode input images by multi-scale feature maps, then decode them to the output that has the same size as input. This process will lose detailed crack information. To tackle this problem, this paper proposed a novel network architecture, Enhanced High-Resolution Semantic Network (EHRS-Net), to maintain and enhance detailed information of the feature maps through convolution procedure, thus, improving the overall crack detection accuracy. The contributions of this paper are: (1) Proposed Resolution Maintain Flow (RMF), which is featured by three different semantic representation extraction flows in parallel with semantic information exchange; (2) Proposed Stacked Atrous Spatial Pyramid Pooling (SASPP) module to enhance the output of the semantic features; (3) Developed a new hybrid loss function to fit proposed network architecture. The proposed methods are evaluated on two pavement crack datasets: an expanded public crack forest dataset (CFD-ex) and a new dataset called HRSD (high-resolution semantic dataset). Comprehensive comparative experiments proved the superiority of the proposed method for pavement crack detection (93.353 % mPA (mean pixel accuracy) and 78.328% mIoU (mean intersection over union) on CFD-ex; 77.159% mIoU on HRSD), especially for tiny cracks and noised pavement cracks.