Automated brake caliper identification via multi-view deep learning for automotive applications

This work deals with the development of a tool for automated identification of brake calipers in the automotive industry. Specifically, we propose multi-view classification of brake calipers using a newly created dataset in which each caliper is captured from three angles, each of 120°. The experimental results show that the proposed method is able to accurately identify brake calipers even in the presence of small variations in the shape of the calipers. We systematically analyzed five representative approaches to multi-view classification: (1) ResNet + LSTM, (2) CNN architecture with late fusion; (3,4) CMAL in both single and multi-view variants to capture the interaction between layers and views; and (5) multi-view Vision Transformers. Experiments show that multi-view modeling leads to consistently good classification results, but individual methods differ significantly in their ability to exploit coherence between views. Attention-based approaches show higher robustness to changes in viewing angle, while recurrent fusion tends to degrade performance when the order of views is ambiguous. The study offers a critical comparison of current methods and provides conclusions for the design of future architectures aimed at the effective use of multi-view data. The entire architecture is designed as a usable solution for real-world automotive needs, from the physical station for capturing images, to the creation of the dataset itself, to the software pipeline for classification. • A multi-view dataset of 65 brake caliper classes captured from three 120° angles was created for industrial classification. • Five deep learning architectures are compared for multi-view brake caliper recognition, including CNN, CMAL, and ViT. • The proposed multi-view Vision Transformer achieves 99.51% accuracy and F1-score of 1.00. • Attention-based models outperform recurrent fusion in exploiting view coherence. • A complete end-to-end pipeline from image capture station to deployed classifier is presented.