Open Set Object Detection in Complex Petrochemical Scenes

With the rapid development of large models, the application of robots for inspection tasks leveraging open-set object detection algorithms has become increasingly prevalent. However, challenges in the petrochemical industry such as dense distribution, occlusion, insufficient lighting, and the presence of numerous small targets have led to issues of reduced accuracy and extended processing times for open set object detection and recognition. This study introduces an enhancement to the Grounding-DINO model, named Sevnce-Det, to address these challenges. The enhancement involves replacing the text encoding component with an offline version. Additionally, it incorporates an attention mechanism, substitutes the image encoding network's backbone with a lightweight architecture, and modifies the feature fusion layer. The effectiveness of the proposed methodology is validated via both ablation studies and comparative analyses, affirming its substantial improvements. The enhanced network attained a mean Average Precision (mAP) of 87.5% on the SevnceData petrochemical dataset.