MSPNet-based Liquid medium leakage Detection in Industrial Sites

Liquid medium leakage is a common issue at industrial sites, posing significant threats to production safety, environmental pollution, and health risks if not promptly addressed. Detecting and warning about liquid medium leaks quickly and effectively is crucial for ensuring enterprise and environmental safety. The traditional object detection background is complex and diverse, but liquid medium leakage detection typically features a more uniform background affected by factors like illumination, leading to background overfitting and decreased detection accuracy. To enhance robustness and eliminate noise interference, we introduce the EMA image enhancement module in the Neck part of MSPNet. This module ensures spatial semantic features are evenly distributed across feature groups, enhancing the algorithm’s image feature extraction capabilities. Additionally, the introduction of the Bottleneck Transformers module significantly improves baseline object detection accuracy while reducing parameter count. During the inference stage, Gaussian modeling and thresholding methods are employed for background modeling to eliminate irrelevant background information, facilitating object detection in images. Experimental results demonstrate a 12.62% increase in average accuracy compared to the baseline, enhancing liquid media leakage detection accuracy under noise interference and mitigating false alarms and omissions.