DSNet: A Novel Convolutional Neural Network Based on Improved Encoder-Decoder for Crack Segmentation of Asphalt Pavement

To address the significant scale differences between cracks and backgrounds and the loss of spatial detail in segmentation processes, which subsequently impacts segmentation accuracy, a dual-scale cross attention and supervised decoding network (DSNet) based on the encoder-decoder was put forward. In the Encoder phase, the ability of the to identify small scale targets was enhanced by processing pavement crack images at different scales. In the Decoder phase, a supervised decoding module was added to adjust the weights of targets at different scales. Additionally, the inclusion of spatial information from down sampling structures increased the learning capability during decoding, thereby improving the effectiveness of crack segmentation. The results demonstrate that the DSNet exhibits better performance than UNet, SegNet and DeepLabV3+, with F1-Score and MIoU values reaching 0.8405 and 0.7644 on the CFD dataset, 0.9033 and 0.8309 on the Crack500 dataset, and 0.8172 and 0.7295 on the CHD dataset, respectively.