EG-Net for Infrared Gas Leak Segmentation under Complex Backgrounds

To address the challenges of low contrast, blurry boundaries, and class imbalance of infrared gas leak targets in complex industrial scenes, this paper proposes a collaborative optimization network termed EG-Net (EfficientNet-based Gas Network) for refined infrared gas segmentation. The method employs EfficientNet-B4 as the encoder to strengthen feature extraction under low signal-to-noise ratio conditions, and integrates the nested dense skip connections of U-Net++ to achieve multi-scale semantic information fusion, thereby improving segmentation performance for small targets and diffuse boundary regions. To cope with the extremely low pixel proportion of leak regions, a Dice-Focal hybrid loss function is designed to mitigate the training bias caused by class imbalance. Furthermore, a post-processing strategy composed of morphological closing and connected-component area filtering is introduced to suppress false-positive noise in complex thermal radiation backgrounds. On the public Gas-DB dataset, EG-Net is compared with main-stream models including U-Net, DeepLabV3+, YOLOv8m, and RT-CAN. Experimental results demonstrate that EG-Net achieves the best overall performance, with a mean Intersection over Union (mIoU) of 59.63% and a Dice coefficient of 68.69%, and attains an image-level accuracy of 92.66% on this core industrial metric. The proposed method effectively improves the accuracy and stability of infrared gas leak segmentation in complex industrial environments, offering reliable technical support for industrial safety inspection and intelligent early warning.