Material Classification in Acoustic NLOS Environments Using an Attention-Based U-Net and Multimodal Fusion With the ANLOS-R Dataset

Non-Line-of-Sight (NLOS) object detection is an important research problem for inferring position, shape, and material properties in situations where direct optical access is not possible. This study proposes a multi-modal learning architecture with attention mechanisms to classify material types using only reflected acoustic waves. The proposed system consists of an attention-supported U-Net-based Reflection Isolation Network, which isolates reflection regions, and a multi-modal fusion classifier that combines spectral-temporal features. Within the scope of the study, the ANLOS-R (Acoustic Non-Line-of-Sight with Reflection) dataset was created, consisting of 1,440 echo samples collected using three different microphone–speaker configurations. This dataset was used for attention-based segmentation of echo intensity maps and extraction of acoustic features; the obtained spectral and temporal representations were evaluated using both deep learning and machine learning-based classifiers. Experimental results show that the single feature-based GRU (Gated Recurrent Unit) model achieved 65% accuracy, while the proposed fusion model reached 74% accuracy. Furthermore, the system demonstrated high generalization capability in both single and multiple material combinations. The proposed method improves NLOS detection in acoustic environments with low signal-to-noise ratios and provides a robust foundation for future acoustic imaging and autonomous detection systems.