An interactive cross-multi-feature fusion approach for salient object detection in crack segmentation

Abstract

Salient object detection (SOD) is a crucial preprocessing technique in visual computing, which identifies the salient regions in an image by simulating the human visual perception system. It achieves remarkable results in tasks such as image quality assessment, editing, and object recognition. However, due to the particularity of pavement crack detection in terms of scale and feature requirements, the SOD model is rarely applied in pavement surface crack detection at present. In order to break the existing dilemma, this paper proposes a new SOD model (iU2Net) specialized for crack detection, which is based on the encoder–decoder structure of U2Net and incorporates the developed interactive cross-multi-feature fusion module (ICMFM). Compared with the existing models, the main contributions of iU2Net are reflected in two aspects. On the one hand, current models are difficult to comprehensively extract the complex features of cracks while iU2Net achieves a breakthrough in feature extraction by efficiently aggregating multiscale crack features and accurately reconstructing them through its unique architecture. On the other hand, iU2Net focuses on infrastructure crack detection, breaking the limitation of independent processing of traditional feature channels and facilitating information exchange. To validate the model's effectiveness, comprehensive experiments are conducted on a public benchmark dataset. iU2Net is compared with eight existing SOD models (EGNet, PoolNet, MINet, F3Net, U2Net, SegNet, BASNet, and DeepCrack). Training and detection performance is evaluated using average mean absolute error (AveMAE), maximum F1 score (MaxF1), mean F1 score (MeanF1), precision–recall curves, and visualizations. Experimental the results indicate that iU2Net exceeds the behavior of other networks during both the training and testing phases, with MaxF1 and MeanF1 achieving maximum values of 0.912 and 0.730, respectively; and AveMAE of 0.048, which is only 0.005 higher than the minimum value, which demonstrates its effectiveness for pavement surface crack detection and indicating potential future applications involving interactive feature fusion.