Data‐Informed SSD Object Detection for Real‐Time Industrial Applications

ABSTRACT This study investigates dataset‐adaptive optimisation strategies for a Single Shot MultiBox Detector (SSD) network, focusing on a critical case study of parts detection on a manufacturing line. We address the challenges of achieving high‐performance object detection in real‐time industrial environments by implementing a structural optimisation approach that dynamically reduces feature maps and prior boxes based on dataset characteristics. The network self‐identifies and prunes redundant features by analysing the statistical properties of the images and objects in the datasets, leading to a leaner and more efficient SSD network. This dataset‐informed reduction of feature maps and anchor boxes accelerates training and inference speeds and enhances detection precision, which is crucial for accurate part identification in high‐throughput production. Experimental results on a custom manufacturing line parts dataset, consisting of multi‐coloured and differently sized bolts and nuts, demonstrate a mean Average Precision (mAP@50) of approximately 96%‐99%, while the inference time was reduced by more than four times (to approximately 3–8 ms in the study case) according to feature extraction backbones, compared to the original network.