Weighted adaptive active transfer learning for imbalanced multi-object classification in construction site imagery

Construction site monitoring relies on robust image classification to enhance safety, track progress, and optimize resource management. However, the amount of clutter and the high cost of manual labeling pose significant challenges. This paper presents an approach to multi-object classification in construction sites using Adaptive Active Transfer Learning. The Weighted Active Transfer Learning with Adaptive Sampling (WATLAS) framework is introduced, where Transfer Learning is combined with weighted Active Learning to efficiently classify diverse objects. A pre-trained InceptionV3 architecture integrated with bidirectional long short-term memory (BiLSTM) layers is utilized, and superior performance is achieved through adaptive sampling techniques compared to traditional methods. WATLAS achieves 97 % accuracy on a comprehensive dataset of 9344 construction site images spanning 15 object categories and maintaining 90 % accuracy with only 5 % labeled data. By optimizing performance metrics, the framework demonstrates significant improvements over traditional methods, making it a scalable solution for construction site monitoring. • WATLAS achieves 97 % accuracy on 15 construction object classes with the full dataset. • With only 5 % labeled data, WATLAS maintains 90 % accuracy, minimizing manual labeling needs. • Integrates InceptionV3 and BiLSTM, outperforming other models in construction object classification. • Adaptive class weighting in WATLAS boosts detection of rare and underrepresented construction objects. • First formal research to extend and implement the ATLAS framework with weighted adaptive sampling. • Excels at classifying rare elements like Circular Ducts (179 samples) and Drywall Panels (54 samples). • Demonstrates robust performance on imbalanced datasets and diverse real-world construction imagery.