Underwater object detection provides strong technical support for underwater exploration, marine scientific research, marine military applications, and cultural conservation. However, efficient and accurate underwater object detection is still facing great challenges due to dim underwater light and image distortion. This paper proposes an improved YOLOv7 with Bi-level Routing Attention (BRA). Specifically, we use Contrast Limited Adaptive Histogram Equalization (CLAHE), sharpen, emboss, and random-brightness-contrast to enhance an underwater image. We further introduce BRA to accurately characterize the targets. Then, we use Focal-EIOU loss to solve the issue of unbalanced samples in the underwater datasets. By doing so, the model focuses more on high-quality anchor boxes, resulting in quick training convergence. Our experimental results show that the detection accuracy on the URPC underwater dataset is improved from 79.3% to 85.6%, which is 6.3% higher than that of the common YOLOv7.