Bubbles in concrete are a critical factor influencing the material’s frost resistance and durability. Detecting these bubbles is essential for assessing concrete quality, a process that holds substantial value for research. Traditionally, concrete bubble detection has relied on manual methods, which are not only cumbersome but also inefficient. To address these limitations, this paper introduces BD-YOLO, the first automatic concrete bubble detector developed using the one-stage object detector YOLOv7. We have designed an enhanced feature pyramid structure, the Effective Feature Pyramid Network (E-FPN), which integrates multi-scale information and leverages the initial layer features of the backbone network to improve detection accuracy. Additionally, we have developed a lightweight version of the Efficient Layer Aggregation Network (ELAN), incorporating Spatial and Channel Reconstruction Convolution (SCConv) to reduce model parameters and suppress feature redundancy within the YOLOv7 architecture. A key contribution of this work is the creation of the first concrete bubble detection dataset, facilitating the training of deep learning models and advancing research in this domain. Comparative experiments with state-of-the-art models reveal that BD-YOLO has achieved a 9.8% increase in precision and a 1.4% increase in recall over the advanced YOLOv7, while also reducing the model’s parameters and floating-point operations by 2.1M and 2.1G, respectively. The source code for BD-YOLO is publicly available at https://github.com/CV-Det/BD-YOLO .