基于改进ResNet50的天然气泄漏红外图像检测方法OACHSSCD

Aims:Aiming at the challenges in infrared image detection of natural gas leakage,such as high similarity between leakage regions and background,variable morphologies,and environmental interference,an image detection method based on improved ResNet50 was proposed.Methods:Firstly,an improved multi-branch RepVGG module was used to reconstruct the feature extraction network.The multi-branch structure was adopted for training to enhance feature expression;and structural reparameterization was applied to convert it into a single-branch structure during inference to maintain high computational efficiency.Secondly,a ConVit module was constructed;and a multi-head self-attention mechanism was introduced into the backbone network to strengthen the model's ability to model the spatial relationship between leakage regions and the background.Finally,the fully connected output layer of the baseline ResNet50 model was replaced with a convolutional layer to reduce the number of parameters and computational complexity and to improve the generalization ability.Results:Test results on the GasVid and IOD-Video datasets showed that the detection accuracies of the proposed method reached 99.83％and 99.80％,which were 1.32％and 1.85％higher than those of the baseline model,respectively.Conclusions:The proposed method significantly improves the feature representation ability in complex scenes and can realize accurate detection of natural gas leakage regions.