SEB-YOLO is an improved YOLOv5 model designed for small target detection in remote sensing images. The model enhances the backbone network with a SPD-Conv module to retain global features and reduce feature loss. A pooling module with an attention mechanism is added to improve target identification. A bidirectional feature pyramid network (Bi-FPN) with bilinear interpolation upsampling is introduced to enhance cross-scale feature fusion. The decoupled head is used to separate classification and regression tasks, improving model convergence and detection performance. Experimental results on the NWPU VHR-10 and RSOD datasets show that SEB-YOLO achieves mAP values of 93.5% and 93.9%, which are 4.0% and 5.3% higher than the original YOLOv5l. The model demonstrates superior performance in detecting small targets in complex remote sensing images. The improved model also shows better results on the RSOD dataset, achieving an mAP of 93.9, which is higher than other models. The algorithm effectively reduces false detection and missed detection rates, and improves detection accuracy and efficiency. The model's performance is validated through ablation studies and experiments on various target categories, showing that it outperforms existing methods in small object detection. The proposed method enhances the ability to capture global context and improves the detection of small targets in remote sensing images.SEB-YOLO is an improved YOLOv5 model designed for small target detection in remote sensing images. The model enhances the backbone network with a SPD-Conv module to retain global features and reduce feature loss. A pooling module with an attention mechanism is added to improve target identification. A bidirectional feature pyramid network (Bi-FPN) with bilinear interpolation upsampling is introduced to enhance cross-scale feature fusion. The decoupled head is used to separate classification and regression tasks, improving model convergence and detection performance. Experimental results on the NWPU VHR-10 and RSOD datasets show that SEB-YOLO achieves mAP values of 93.5% and 93.9%, which are 4.0% and 5.3% higher than the original YOLOv5l. The model demonstrates superior performance in detecting small targets in complex remote sensing images. The improved model also shows better results on the RSOD dataset, achieving an mAP of 93.9, which is higher than other models. The algorithm effectively reduces false detection and missed detection rates, and improves detection accuracy and efficiency. The model's performance is validated through ablation studies and experiments on various target categories, showing that it outperforms existing methods in small object detection. The proposed method enhances the ability to capture global context and improves the detection of small targets in remote sensing images.