An improved YOLOv8 corrosion bolt detection model was proposed to address the challenges in detecting corroded bolts in shield tunnel scanning images, particularly the issues of small target sizes, occlusion, and large image dimensions. A focal modulation module was introduced in the backbone, and the neck was upgraded to a weighted bidirectional feature pyramid network (BiFPN) with four levels of feature fusion, to enhance feature extraction ability for small targets. A separated and enhancement attention module (SEAM) was incorporated into the detection head, effectively addressing the problem of bolt occlusion. Slicing aided hyper inference (SAHI) was applied during the detection phase, allowing the model to process large-sized scanned images efficiently. A rust grade determination method based on HSV (hue, saturation, value) color space segmentation was proposed. The color contrast between the rusted areas and the surrounding background was utilized to accurately segment the rust regions, classify the corrosion severity, and overlay the results directly onto the scanned image. Experimental results showed that the improved YOLOv8 corrosion bolt detection model achieved a mean average precision of 95.0%, with a compact model size of 3.9 MB and a frame rate of 83.3 frames per second, enabling high-performance detection of corroded bolts.