Soft errors pose a critical reliability threat to deep neural networks in safety-critical systems. We present YOLO-FI, a fault injection framework for YOLO models that supports module-wise analysis, multiple formats, and customizable fault models. Using YOLO-FI, we evaluate YOLOv8 in floating- and fixed-point forms on COCO. Results show that floating-point YOLOv8 is vulnerable across modules, while fixed-point models are resilient except for the final module. Range restriction improves resilience in most floating-point modules but fails in the final one. To address this, we propose selective hardening of the vulnerable module, achieving effective fault mitigation with modest overhead.