| description abstract | This paper introduces DLANet, an algorithm for semantic segmentation designed to enhance the pixel-level detection of sealing cracks. DLANet extends the DeepLabV3+ encoder-decoder architecture by integrating four different scales of feature maps in the encoder. This facilitates a wider exchange of information, thereby enhancing the model’s performance. Moreover, DLANet integrates multiple fusion techniques, attentional mechanisms, and self-attention to more effectively capture features associated with sealed cracks. The data set, comprising 4,658 images, is partitioned into training, validation, and test sets. The test set comprises 1,382 image sets. On the test images, DLANet achieved an F-measure of 81.62% and an intersection over unity rate of 73.31%. Experimental findings demonstrate that DLANet surpasses five state-of-the-art semantic segmentation models in terms of detection accuracy. Sealed cracks are a type of surface defect on roads. Damage to sealed cracks can affect road aesthetics, vehicle travel, and comfort. Water on the road surface can infiltrate through damaged sealed cracks, causing structural damage to the road, shortening its service life, and significantly increasing road maintenance difficulty. Although extensive research has been invested in the intelligent detection of sealed cracks, the results have not yet met optimal expectations. This study proposes a new semantic segmentation network, DLANet. When compared with multiple classic networks on public and proprietary data sets, DLANet demonstrates significant improvements in recognition accuracy. Furthermore, through the application of the network model in engineering practice, the detection results are highly accurate, thus confirming the robustness and reliability of the proposed DLANet. | |
