The primary challenge in detecting small objects stems from their inherently limited size, necessitating strategies such as enlarging the size of input images or feature maps. In this regard, researchers often employ super-resolution (SR) techniques. While existing methods can partially address this issue, several bottlenecks remain: 1) It is challenging to mitigate the high computational overhead associated with high-resolution (HR) images or feature maps while enhancing detection performance. 2) There is a distinct separation between the SR network and the detection network in terms of architecture design. To overcome these, we first establish a design concept of an SR network tailored for small object detection (SOD) and propose the SR Auxiliary Branch for SOD (SOD-SRB) with bidirectional semantic information matching. We use ablation experiments to validate the rationality of our method. Comparative experimental results indicate that SOD-SRB outperforms the state-of-the-art (SOTA) SR auxiliary branch designed for SOD and can enhance the detection performance of baseline across various detectors on different small object datasets.