Enhancing Object Detection in Layout Analysis: Leveraging Vision Transformer and Fourier Neural Operator

This paper investigates the application of Vision Transformer (ViT) and Fourier Neural Operator (FNO) models in layout analysis tasks, aiming to enhance object detection accuracy and robustness in diverse document layouts. Traditional convolutional neural network (CNN) models often exhibit limitations in handling non-standard document layouts, leading to detection errors. The proposed FNO-YOLO model integrates FNO modules into the VitT-YOLO backbone, demonstrating superior detection performance and reduced computational complexity compared to Self-Attention-based models. Experimental results on the CDLA dataset and engineering drawings showcase the efficacy of FNO in maintaining robust object detection performance, especially in out-of-sample evaluations. FNO's resolution-agnostic characteristics and incorporation of physical mechanisms contribute to its generalization capability and resilience to overfitting, making it a promising solution for layout analysis tasks in document digitization and information management.