To address the issues of low defect detection accuracy and poor real-time performance in the internal inspection of metal bellows, an improved algorithm based on YOLOv8n, named YOLOv8n-FCW, was proposed. First, the backbone network of YOLOv8n was replaced with the faster and more lightweight faster neural networks (FasterNet), ensuring a balance between detection accuracy and computational efficiency. Second, a C2F_S module, incorporating the channel attention mechanism squeeze-and-excitation layer V2(SeLayerV2), partial convolution (PConv), and pointwise convolution (PWConv), was utilized to enhance the network capability for global feature learning, thereby improving the precision of feature representation and the integration of global information. Finally, a loss function named weighted IoU_Scale (WIoU_Scale), which combines weighting and scale adjustment, was adopted to accelerate the convergence speed and improve the accuracy of anchor box predictions. Experimental results demonstrate that, compared to the original YOLOv8n model, YOLOv8n-FCW achieves improvements of 4.7%, 3.8%, and 4.1% in mAP50, mAP50-95, and F1-Score, respectively. The model contains 7.3 million parameters and requires 11.2 GFLOPs of computation. Compared with other object detection algorithms, YOLOv8n-FCW exhibits higher accuracy and parameter efficiency, effectively meeting the requirements for internal defect detection in metal bellows.