Accurately and quickly obtaining information from garbage bins has great application value in smart city construction and urban environmental management. However, existing deep learning methods are affected by factors such as occlusion, large geometric appearance differences, and multi-scale, leading to missed detections in garbage bin detection results. We propose a Cot-DCN-YOLO model for garbage bin detection, which is designed to effectively extract contextual information with the Double Convolutions Semantic Transformation (DCST) module, which addresses the vulnerability of garbage bins to occlusion. According to the large geometric appearance differences when garbage bins are damaged, we propose the C2f embedded with DCNv2 (DC2f) module, which can adaptively adjust the target shape with a flexible receptive field. Furthermore, considering the multi-scale characteristics of garbage bins in images, we introduce the SPPCSPC module. Experimental results show that compared with other methods, Cot-DCN-YOLO achieves the best results on our self-made garbage bin dataset, with Precision, Recall, and mAP reaching 77.1%, 69.4%, and 74.0%, respectively, outperforming existing SOTA methods.