Enhancing Object Detection Accuracy in Vision Transformers Through Architectural Optimization and Structured Parallelism

Jarata A.A. · Proceedings - International Conference of the Chilean Computer Science Society, SCCC (2025) · DOI: 10.1109/sccc67219.2025.11420732

Resumen

Object detection remains a core task in computer vision, demanding models that are both accurate and computationally efficient. This work presents a precision-optimized Vision Transformer (ViT) architecture designed to improve object detection performance through two key strategies: architectural refinement and structured parallelism. The proposed model incorporates enhancements in positional encoding, regularization, and attention mechanisms to increase detection accuracy, while also leveraging pipeline and farm parallelism patterns to accelerate training and inference. Evaluated on the Caltech 101 dataset with a focus on aircraft detection, the model achieves a high Intersection over Union (IoU) of 0.92 and demonstrates strong generalization performance. These results highlight the effectiveness of combining model-level improvements with task-parallel execution to achieve both high precision and efficiency in real-time object detection scenarios.

Instituciones: Universidad Nacional de San Agustín de Arequipa