CSF-Net: Cross-Stage Fusion Network with Dual Backbone for Small Object Detection

Small object detection remains a challenge in computer vision due to low pixel occupancy, feature scarcity, and susceptibility to background interference. Conventional single-backbone networks often struggle to balance deep semantic extraction with the preservation of shallow details. Deep down-sampling can lead to the loss of edge and texture information, while late-stage fusion may fail to recover these details effectively. To address these limitations, this paper proposes a Cross-Stage Fusion Network with a Dual Backbone (CSF-Net). Our network employs an asymmetric design: a shallow backbone maintains a higher resolution to preserve fine-grained details, while a deep backbone extracts contextual semantics. These two streams interact via progressive cross-stage connections, facilitating the early fusion of small object information. Experiments on the Micro RGB UAV dataset indicate that CSF-Net improves the mAP of the YOLOV8 baseline from 62.8% to 67.0%, validating the effectiveness of the proposed architecture in enhancing detection performance for small targets.