To enhance the accuracy of detecting apparent defects in concrete bridges, this study proposes an improved model based on the YOLOR. Initially, a concrete bridges defect (CBD) dataset comprising 7009 images was established, encompassing nine distinct types of defects: pockmark, spalling, cavity, whitening, reticular crack, single crack, rust expansion, exposed rebar and leakage. Subsequently, using the CBD dataset, performance comparison experiments involving YOLOv3, YOLOv4, YOLOR-640, and YOLOR-1280 were conducted to select the optimal model. Based on the characteristics of the apparent defects in bridges, the anchor boxes and localization loss function were optimized respectively, thereby proposing an improved model named YOLOR-BDIM. The results indicate that the mAP of YOLOR-BDIM increased by 2.7 % compared with initial model. Furthermore, YOLOR-BDIM achieved a detection accuracy of 97.5 % during the random tests, showing excellent detection performance even in complex environments.