Small construction materials detection: an approach of enhanced feature extraction and representation

Automated construction materials detection is crucial for material lean management, such as material planning, inventory, site usage, and monitoring. However, there are numerous small materials in the construction site due to the long monitoring distances, which easily cause missed and incorrect detection owing to their indistinguishable features and complicated backgrounds. To improve detection accuracy for small materials, this study proposes an augmented detection method based on enhanced feature extraction and representation. In the proposed method, DenseNet is utilized as the backbone to enhance the feature extraction of small materials. Additionally, the explicit visual center is introduced to enhance the feature learning of small materials. Finally, the multi-scale detection structure is optimized by adding a scale to improve feature representation. Experimental results demonstrate that the average precision for small objects (APs) have improved by 5.3%, and the mean average precision (mAP) has reached 84.3%, surpassing other state-of-the-art methods. The proposed method also exhibits strong adaptability to various conditions such as shadows, blurriness, and cluttered backgrounds. Additionally, the impacts of different backbone networks and detection scales on accuracy are discussed. This research provides theoretical and practical references for material lean management and facilitates the application of digital twin in materials management.