A method of hierarchical feature fusion and adaptive receptive field for concrete pavement crack detection.

Automatic crack detection is a key task to ensure the quality of concrete pavement and improve the efficiency of pavement maintenance. To address the problem of failing to adaptively combine multi-scale spatial information and the loss of crack detail information in crack detection, a network model with hierarchical feature fusion and adaptive receptive field has been proposed. Firstly, the improved SKNet serves as the backbone network for extracting multi-scale features. Subsequently, the corresponding attention mechanism is introduced to optimize the side output, enhancing attention to the crack location and channel information. Finally, we propose a method that fuses spatial separable convolution and attention mechanism, and design a spatial attention fusion module to restore more crack details. The side network integrates low-level features and high-level features at multiple levels to assist in obtaining the final prediction map. To verify the validity of the proposed method, we evaluate it on three publicly available crack datasets: DeepCrack, CFD and Crack500, achieving F-score ( F 1 ) values of 87.20%, 63.53% and 62.26%, respectively. [ABSTRACT FROM AUTHOR]