<p>This study presents the FSO-YOLO network to address the challenges of diverse morphological variations and small object misdetection in large-volume parenteral (LVP) solution inspection. The model incorporates multiple specialized optimizations to enhance detection performance, specifically for foreign object detection in parenteral solutions. We propose a novel input layer structure, the high-resolution lightweight fusion module (HRLF), designed to address the inherent contradiction between fixed downsampling strategies and small object feature preservation in lightweight object detection networks. Furthermore, we introduce a global cross-attention (GCA) module that simultaneously considers both channel information and spatial position attention, thereby improving the model’s ability to focus on small objects. The model also combines deformable convolutional networks (DCN) with generalized efficient layer aggregation networks (GELAN) to enhance feature extraction capabilities, improving detection effectiveness across various types of contaminants. Experiments conducted on our LVPD dataset demonstrate that the proposed method achieves approximately 8% improvement in mean average precision (mAP) compared to baseline networks while maintaining excellent real-time performance.</p>