Enhanced YOLOv8 Algorithm for Real-Time Instance Segmentation of Components in Modular Integrated Construction

Detecting module components at the factory is crucial for safety monitoring, quality control, and productivity enhancement. However, traditional segmentation methods are neither cost-effective nor capable of achieving real-time performance. To address these challenges, this study proposes an improved YOLOv8 modular integrated construction segmentation algorithm. The proposed method introduces the construction of a small object-YOLO, optimizing the YOLOv8 model by replacing the basic module with a novel cross-stage partial network fusion module. This new module employs deformable convolutional networks v2 to manage geometric variations of objects and focus on relevant image regions. Additionally, the Wise-IoU strategy reduces the competitiveness of highquality anchor boxes and mitigates harmful gradients generated by low-quality examples. The MultiHead self-attention mechanism further enhances detection accuracy by capturing the relationship between the image and significant objects, making it more suitable for the modular integrated construction dataset. Given that construction images are often taken from a top or bird's-eye view, small objects can be challenging to be detected. Therefore, this algorithm incorporates a small object detection algorithm to improve the model's capability in identifying small objects. Experimental results demonstrate that the improved YOLOv8 model effectively identifies moving objects, achieving a 4.4% increase in mAP and a 4.3% increase in F1 score compared to the original YOLOv8 model, while reducing parameters by 54.05% and GFLOPs by 55.39%. The proposed algorithm provides a reference for automatic segmentation methods of modular integrated construction components at the factory.