Enhanced Square Fiducial Marker Recognition under Challenging Visual Environments Using Multi-Scale CNN-Transformer Fusion

Received: 11 June 2025 | Revised: 11 July 2025, 23 July 2025, and 4 August 2025 | Accepted: 11 August 2025 | Online: 2 September 2025

Corresponding author: Liliek Triyono

Abstract

Recent methods using deep learning have demonstrated promising outcomes in tackling the issue of object recognition in low-light images. However, existing techniques often face challenges related to distortion and occlusions, and many strategies rely on neural networks with convolutional neural network (CNN) structures, which are limited in their ability to capture long-term dependencies. This frequently leads to inadequate recovery of very dark areas in low-light images. This work introduces a unique Transformer-based method for ArUco marker recognition in low-light environments, termed Extreme ArUco Vision Transformer (XAViT). We present a Transformer–CNN hybrid block that utilizes mixed attention to effectively capture both global and local information. This method integrates the Transformer's capacity to model long-range dependencies with the CNN's proficiency in extracting detailed features, facilitating the reliable detection of ArUco markers even in extreme lighting conditions. Additionally, we employ a Swin-Transformer discriminator to selectively improve various areas of low-light images, alleviating problems of overexposure, underexposure, and noise. Comprehensive experiments show that XAViT achieves 99.16% accuracy, 97.86% recall, 97.95% precision, and 97.89% F1-score on a realistic low-light dataset, outperforming state-of-the-art CNN and Transformer models. Moreover, its utilization in additional vision-based tasks underscores its potential for wider implementation in advanced vision applications.