Accurate detection of corner castings is pivotal for enhancing the performance of automated crane systems for containers, to streamline the loading and unloading operations of ships. Bounding box regression, a fundamental component of object detection, plays a critical role in precisely identifying and localizing these corner castings. In this study, we have developed an algorithmic framework aimed at integrating annotation outputs from both single-shot detection (SSD) and zero-shot detection (ZSD) methodologies. We introduce a novel technique for computing deviations in the spatial centrality of bounding boxes relative to their respective detected objects. Our approach enhances bounding box regression, resulting in significantly improved Best-Fit Bounding Box (BFBB). Our findings highlight substantial reductions in bounding box size ranging from 17.72% to 36.47%, and centreness accuracy improved, ranging from 7.49 pixels to 25.17 pixels. This demonstrates enhanced BFBB performance due to better generalization of the model with ZSD annotation, underscoring our contribution to advancing automated container port operations.