Pixel-Level Detection of Cracks Based on Loop Semantic Diffusion Integration

Precise crack identification on intricate asphalt pavements poses a significant challenge for intelligent pavement distress detection. This paper proposes the CellNet, a deep learning model designed to offer an effective solution to this challenge. The CellNet integrates global contextual semantic information across multiple loops to improve feature extraction. More specifically, information at various resolutions is iteratively acquired and processed through loops to facilitate semantic diffusion and integration. Moreover, the atrous spatial pyramid pooling (ASPP) module is incorporated at the bottom of each loop structure to improve the extraction of deep features. To prevent long-distance information forgetting within the same resolution level between loop rounds, this paper introduces the transformer and convolution attention module (TACM), which combines convolution and transformer. Experimental results demonstrate that the proposed CellNet achieves an F-measure of 91.89% and an IOU of 85.01% on 1,400 test images. The performance evaluation on both private and public data sets indicates that compared to state-of-the-art semantic segmentation models, the proposed CellNet not only achieves higher detection accuracy but also demonstrates a considerably notable recognition speed. In practical engineering applications, the proposed CellNet also demonstrates good performance.