Bidirectional cross-scale detail-semantic collaborative network (BCDS-Net) for vision-based waste recognition

The continuous growth of global waste generation poses serious challenges to environmental protection, public health, and resource recovery. In intelligent sorting systems, vision-sensor-based waste recognition is a challenging core perception task, yet its reliability is often affected by scale variation, background clutter, and inter-class visual similarity. To address these challenges, we propose BCDS-Net, a Bidirectional Cross-Scale Detail-Semantic Collaborative Network combining a Hybrid Conv-MLP backbone, Scale-Aware Context Gated Fusion, Dilated Scanning Attention, and Bidirectional Cross-Scale Aggregation to enhance local details, model long-range dependencies, and fuse multi-scale features. This study focuses on single-label waste image classification. Each input image is assumed to contain one predominant waste category, and the model outputs one image-level category label. BCDS-Net does not perform object localization, multi-object detection, or instance segmentation. Evaluations on TrashNet, RealWaste, and TrashBox achieved accuracies of 96.73%, 92.15%, and 96.23% and recall rates of 96.87%, 92.26%, and 96.27%, respectively, outperforming representative baselines. Scale- and condition-stratified evaluations further demonstrate stable recognition under object-size variation, background clutter, occlusion, and degraded image quality. With 11.53M parameters and 2.52G FLOPs, BCDS-Net achieves a batch-size-one FP32 latency of 24.4 ms (41 samples/s) on an NVIDIA Jetson Xavier NX at an input resolution of 224 $$\times $$ 224, and 3.2 ms on an NVIDIA RTX 3090. These results indicate that BCDS-Net provides a favorable accuracy–efficiency trade-off and has potential for edge deployment, although performance on a specific device remains hardware and implementation dependent.