To address the challenges in garbage detection in complex urban environments, particularly the difficulties caused by small target sizes and cluttered backgrounds, an improved algorithm named YOLO-SPRL is proposed for small target garbage detection based on YOLOv8. First, the SPD-Conv module is incorporated and a P2 detection layer is added while the P5 detection layer is removed to enhance the detection capability for small garbage objects. Second, a rectangular self-calibration module (RCM) is integrated into the P3 layer of the feature pyramid to suppress background noise and eliminate interference from complex backgrounds. Finally, a learnable separable kernel attention (LSKA) module is introduced in the model’s neck, which reduces computational complexity while strengthening deep feature fusion capability. This optimization reduces model size while maintaining improved accuracy, laying the theoretical groundwork for deployment on embedded devices. Experiments on the VisDrone and custom street view garbage datasets demonstrate that YOLO-SPRL achieves an mAP@0.5 of 38.91%, representing an improvement of 4.22% compared to the baseline model. Additionally, the algorithm exhibits excellent robustness and strong small target recognition capability on a custom drone garbage dataset. The integrated improvement strategy proposed in this study effectively enhances garbage detection performance in complex scenarios. YOLO-SPRL achieves an optimal balance among accuracy, speed, and model size, demonstrating the feasibility for deployment on embedded mobile devices, and providing a reliable technical solution for real-time monitoring and management in urban intelligent sanitation systems.