EM-DETR: Efficient Multiscale DETR for Gas Leak Detection Based on Infrared Imaging

Industrial gases play a crucial role in modern industrial production. However, many industrial gases are only visible in the infrared spectrum. Therefore, using infrared camera for gas leak detection is essential for the normal operation of production lines and the safety of personnel. Existing automatic leak detection algorithms based on convolutional neural networks often involve numerous handcrafted feature components, resulting in complex model architectures, with hyperparameter selection significantly impacting model performance. To address these issues, we propose an efficient and high-performance end-to-end object detection network, the Efficient Multi-scale Detection Transformer (EM-DETR), for automatic gas leak detection. To reduce model computational overhead, we design a new efficient encoder. Additionally, to handle leaks of varying ranges due to gas dispersion, we implement bidirectional fusion of multi-scale features to capture granularity information of gases at different scales. Finally, based on the similar unified distribution of gas leak images, we introduce a pre-auxiliary loss to adjust the learnable object queries initialized. Experimental results demonstrate that the EM-DETR R50/R101 outperforms previous methods, achieving 40.7%/43.2% AP on gas leak infrared image datasets with minimal parameters and computational resources, providing a superior solution for automatic gas leak detection using infrared camera. Code is available at https://github.com/zhushiL/EM-DETR.