Aiming at the problems of low detection accuracy and insufficient generalization ability caused by inconspicuous color features and irregular size in crack images collected by optical sensors in road inspection systems, it is propose improved YOLOv5s crack detection algorithm. The Global Attention Mechanism (GAM) fused with Depthwise Separable Convolution (DSC) is introduced into the backbone feature extraction network to obtain rich cross-dimensional features while reducing the complexity of attention. The recognition ability of cracks is enhanced; the spatial pyramid soft pooling network (Spatial Pyramid Softpool, SPSF) is used to preserve multi-dimensional semantics through Softpool pooling to reduce information dispersion and improve the accuracy of bounding box regression; in the neck feature enhancement network, Downsampling is performed with Atrous Depth Separable Convolution (Atrous DSC), which enhances the aggregation ability of deep and shallow information by expanding the receptive field, and improves the generalization of crack identification. After experiments on the self-made road crack data set, compared with YOLOv5s, the mAP of improved algorithm is increased by 2.2%, which effectively improves the accuracy of road crack detection and the generalization ability of crack recognition under different backgrounds.