Automatic crack defect detection via multiscale feature aggregation and adaptive fusion

In this paper, a multi-scale feature aggregation and adaptive fusion network, is proposed for automatic and accurate pavement crack defect segmentation. Specifically, faced with the linear characteristic of pavement crack defects, a multiple-dimension attention (MDA) module is proposed to effectively capture long-range correlation from three directions, including space, width and height, and help identify the pavement crack defect boundaries. On this basis, a multi-scale skip connection (MSK) module is proposed, which can effectively utilize the feature information from multiple receptive fields to support accurate feature reconstruction in the decoding stage. Furthermore, a multi-scale attention fusion (MSAF) module is proposed to realize effective multi-scale feature representation and aggregation. Finally, an adaptive weight fusion (AWL) module proposed to dynamically fuse the output features across different network layers for accurate multi-scale crack defect segmentation. Experiments indicate that proposed network is superior to other mainstream segmentation networks on pixelwise crack defect detection task.