EFCNet for small object detection in remote sensing images

Object detection, as a crucial component of remote sensing image processing, has become one of the primary methods with the maturation of deep learning technologies. Nonetheless, detecting small objects in remote sensing images remains a significant challenge. Addressing this issue, this study proposes an enhanced network model based on You Only Look Once YOLOv5, aimed at improving the detection capabilities for small objects in remote sensing images. The model employs a novel backbone network, ODCSP-Darknet53, to enhance feature extraction efficiency, and incorporates the small object enhancement bi-directional feature pyramid network (STEBIFPN) structure in the neck region of the network for optimized scaling of small object information. Additionally, we have designed two distinct weighted fusion strategies to further boost the model's performance in detecting small objects. In the detection head portion of the model, a four-head detection network specialized for small objects is constructed, and adaptively spatial feature fusion (ASFF) technology is introduced to optimize the recognition capabilities for small objects. Experiments conducted on the DOTA and DIOR datasets demonstrate that our model achieves an average precision mean of 75.9% and 80.5%, respectively, with the model's parameters and computational requirements amounting to 13.4M and 30.2 GFLOPs, respectively. Compared to the original YOLOv5s model, our model exhibits significant performance improvements in detecting typical small objects such as Bridge and Ship. Thus, this research provides an effective solution for object detection in the field of remote sensing image processing.