Ammonia gas used in industry is colorless, flammable and explosive, and its diffusion is susceptible to interference caused by wind conditions and other meteorological factors. Traditional methods of detecting target locations need to spread the leak to a certain range and contact the sensor to respond, resulting in great safety risks for inspection workers and the environment. Therefore, it is of great significance to find a large-area, efficient, non-contact gas leakage detection method that is in line with the development trend of the times, can effectively solve the potential safety hazards of personnel, and reduce the impact of gas leakage on the environment.This paper proposes a fusion of structure-heavy parametric transformation of the infrared non-contact detection model network model GRNet. The candidate bounding box suitable for infrared detection of gas leakage are analyzed by K-means clustering to preset the model parameters. Whereafter, the gas leak infrared image is preprocessed using the data enhancement method of Mosaic-Gamma transformation, so that the image combines the contextual information of 4 different forms of gas leak areas, enriches the leak scene, and increases the training batch size in disguise during training. This improves the generalization ability of the model and improves detection accuracy. Moreover, the CIoU localization loss function is used to optimize the size and center position of the leakage area to improve the predicted accuracy in locating the leakage area. Finally, the improved lightweight RepVGG module is adopted to reconstruct the feature extraction network instead of the convolutional layer of the feature extraction network, which enhances the feature extraction capability of the model, reduces the number of model parameters, compresses the size of the model, and improves the speed of model inference. The final GRNet model for ammonia leak infrared detection improves the mean average precision, single image test time, model size, and number of parameters by 2.70%, 11.76%, 27.43%, and 28.90% over the original YOLOv5s base model, reaching to be 94.90%, 3.4 ms, 11.30 MB, and 5.47 MB, respectively.Next, this paper adopts the pseudo-color mapping technology to qualitatively analyze the gas leakage concentration to achieve the visual perception effect of the leakage concentration which helps to improve the efficiency and accuracy of the staff's emergency response. And PyQt5 is used as the implementation tool of the graphical system interface to encapsulate the constructed network model, which is more intuitive and easy to operate to achieve the visualization of the interface of the gas leakage infrared detection system. Finally, the effectiveness of the GRNet model for the gas leak detection task is further verified in the embedded development device, the GRNet model in the detection speed detection speed compared to YOLOv3, YOLOv5s improved to reach 3.03 frames/s, while the detection accuracy is consistent with the PC in this paper there is no loss of up to 94.90% accuracy. This indicates that the GRNet model is compatible with faster detection speed while leakage detection is effective, and the feasibility of deployment on an embedded development platform is relatively high.This paper can provide ideas for deep learning model design and leakage concentration analysis as well as deployment for the development of gas leakage non-contact detection devices to ensure the safe production of gas-related enterprises and stable operation.