Research on Symbol Recognition Method of Historical Buildings along the Chinese Eastern Railroad based on Improved YOLOv8s Technology Framework

Abstract

Addressing the prevalent issue of target misdetection within existing algorithmic frameworks for architectural symbol recognition, such as the challenges posed by small-scale targets, compact distributions, and the multifaceted nature of architectural symbol categories, an enhanced YOLOv8s model tailored for architectural symbol recognition is introduced. This model integrates the DCNv3 module within the backbone network architecture, in conjunction with C2f, which augments the model's receptive field, mitigates the attendant increase in model complexity resulting from C2f convolutions, and enhances the model's efficacy in target recognition and convergence. Utilizing the SIoU loss function in lieu of CIOU significantly enhances the training efficiency and inferential accuracy of the object detection model; the incorporation of the D-LKA attention mechanism within the head network further bolsters the detection capabilities for small-scale targets. Experimental findings corroborate that the improved YOLOv8s model achieves an mAP@0.5 score of 85.5% on the Chinese Eastern Railroad historical building symbol dataset, a 3.6% improvement over the baseline YOLOv8s model. Collectively, the refined model markedly elevates the detection prowess for architectural symbol targets, more adeptly fulfilling the requirements of historical building symbol recognition along the Chinese Eastern Railroad.