Object detection of mural images based on improved YOLOv8

Currently, mural object detection is highly dependent on traditional manual detection means, which is inefficient and prone to frescoe damage. Therefore, We propose an enhanced mural image detection algorithm, Brg-YOLO, based on YOLOv8, to achieve efficient, non-contact automatic detection. First, We enhance detection across scales and complex scenes by incorporating a bidirectional feature pyramid network (BiFPN) in the neck, enabling efficient multi-scale feature reuse and improved feature fusion. In addition, we embed the residual squeezing-and-excitation (RSE) attention module in the backbone to mitigate the feature aliasing effect. Finally, with the Ghost+RSE Bottleneck design in the Neck part, we realize a lightweight model deployment that maintains the excellent detection effect while reducing the number of parameters. The experimental results show that the model achieves 84.6% and 47.8% for mAP@0.5 and mAP@0.5:0.95, respectively, in the mural object detection task, which far exceeds similar methods. This study provides new perspectives and tools for mural painting conservation and research, realizes efficient and accurate mural detection through non-contact automatic detection methods, and creates a new paradigm for mural heritage conservation.