Improved YOLO X with Bilateral Attention for Small Object Detection

Currently, general object detection networks exhibit low accuracy and instability in small object scenes. To address this issue, we propose an improved YOLO X algorithm for small object detection based on bilateral attention. This algorithm introduces a bilateral attention module that integrates attention feature maps from both spatial and channel dimensions, adaptively capturing refined features of the target object. This enriches the information of small objects on feature maps of different sizes and enhances communication of potential semantic information. Additionally, a relative position labeling strategy is used to enhance the accuracy of target location information and expand the detection range. Experimental results demonstrate that our method is more effective than the YOLO X model, particularly in handling small objects in remote sensing images, with a 1.1% improvement in mAP0.5.