An improved YOLOv5 algorithm for construction solid waste detection

Construction solid waste contains many small objects, which can limit the model's recognition performance and lead to suboptimal small object detection. This may cause misidentification, missed detection, and repeated detection of construction solid waste, posing challenges for sorting operations. To address this issue, we propose an improved YOLOv5 algorithm for construction solid waste detection. This algorithm incorporates the Vision Transformer (ViT) Transformer Encoder structure into the backbone network and designs a Fusion-Concat module to replace the original Concat module in the Neck network, thus enhancing the model's detection capability. Moreover, we created an annotated construction solid waste detection dataset to validate the model's detection performance. Experimental results show that, while reducing the model's parameter size and computational complexity, the improved YOLOv5 algorithm outperforms the original YOLOv5 algorithm in small target detection, with a Mean Average Precision increase of 1.1%.