<p>Accurate identification of recyclable waste is essential for efficient waste management. However, in real-world scenarios, severe occlusion, geometric deformation, and complex stacking of objects often result in high rates of missed detections and false positives for existing lightweight models. To address these challenges, a high-accuracy and high-efficiency detection model suitable for edge computing was developed to enhance the discriminative capability of complex features. A self-calibrated convolution and convolution–attention fusion lightweight model, termed SCCL-YOLO11n, is proposed. Specifically, self-calibrated convolution (SCConv) modules are integrated into the YOLO11n architecture to adaptively enhance multi-scale feature representations. In addition, a dual-attention module based on deformable attention (C2DA) is introduced to enable the model to focus on non-rigid object contours. This module is followed by a Convolution and Attention Fusion Module (CAFM), which collaboratively captures local details and global contextual dependencies, thereby improving the modeling of deformable targets and the perception of occluded contextual information. Furthermore, a Lightweight Shared Convolutional Detection Head (LSCD) is employed to reduce computational complexity while maintaining high detection accuracy. Experiments conducted on a self-constructed dataset containing 5,383 images demonstrate that the proposed model achieves a mean Average Precision (mAP) of 96.3% with a computational cost of only 6.1 GFLOPs. The model exhibits strong robustness in scenarios involving occlusion, overlap, and deformation, and its overall performance surpasses that of existing mainstream lightweight detection models. This study provides an efficient and reliable edge-based solution for real-time recyclable waste detection in complex environments.</p> Graphical Abstract <p></p>