Deep learning for construction waste detection using ConvNeXt V2 EMA attention and WIoU v3 loss.

With the rapid increase in construction waste, traditional manual sorting methods have become inefficient, and enhancing automatic recognition and classification capabilities has become critical to mitigating resource wastage and environmental pollution. To improve the accuracy of construction waste detection and reduce the false detection rate, this study proposes an optimized detection model, termed YOLO-CEW, designed to enhance both the precision and robustness of construction waste detection. The model integrates the ConvNeXt V2 module, EMA multi-scale attention mechanism, and WIoU v3 loss function, and is developed by optimizing YOLOv8 as the baseline model. These improvements strengthen feature extraction capability, increase localization accuracy, and effectively reduce the interference of low-quality samples. Using the Construction and Demolition Debris Object Detection (CODD) dataset for training and testing, experimental results demonstrate that the proposed model achieves a precision of 96.84%, recall of 95.95%, mAP@50 of 98.13%, and an F1-score of 96.39%, all surpassing the original baseline model. Ablation studies further confirm the substantial contribution of each improved module to overall performance. In addition, statistical testing using the bootstrap method validates that the performance gains of the model are statistically significant. Comparative experiments with multiple YOLO variants show that YOLO-CEW exhibits superior accuracy and recall in construction waste detection tasks and thus holds strong potential for practical applications. This study provides a more efficient technical solution to support smart city development and the management of construction waste.