Enhanced mask R-CNN for luminaire detection through brightness balancing and distribution-guided optimization in tunnels

Luminaire detection methods based on deep learning encounter challenges in tunnel luminaire detection due to the complex environment and unfavorable lighting conditions. To overcome these issues, this paper proposes an improved tunnel luminaire detection solution by enhancing the Mask R-CNN using brightness balancing and distribution-guided optimization. Leveraging tunnel gray-scale images and the Mask R-CNN object detection framework, a feature fusion network based on ResNet-FPN, trained via transfer learning, which enhances performance in detecting object luminaires. Furthermore, considering the differences in luminaire brightness, their backgrounds, and regular spacing, an object brightness enhancement method based on a fixed threshold and an innovative distribution-guided optimization strategy are introduced to effectively reducing missed detections and false alarms while accurately correcting the detected luminaire positions. To evaluate the performance of the proposed approach, real datasets of tunnel environment are used. Experimental results revealed that the proposed approach achieved precision, recall, and an F1-score of 98.3%, 98.5%, and 0.984, respectively, which improved of 0.5%, 5.1%, and 0.029, respectively, compared to the original model, thus, could be applied to the 3D model construction and intelligent management of tunnels. The code implementing our method is available at https://github.com/xubintao/OBE-and-DGO.