【摘要】 非制冷型红外相机由于其成本低、寿命长、性能稳定等优势在气体泄漏检测领域有着广泛应用，而良好的图像去噪算法可以有效提升其检测灵敏度与准确性。结合深度学习和迁移学习技术，提出了一种基于深度迁移学习的气体泄漏红外图像去噪方法。首先使用静止场景数据集对卷积神经网络模型进行训练，然后固定部分模型参数，并通过仿真气体数据集对模型再次训练，最终获得适用于气体泄漏红外图像去噪的模型。实验结果表明，该方法可以对非制冷型红外相机拍摄的气体红外图像进行去噪，去噪后的图像具有明显的气体轮廓信息，同时可以分辨出泄漏源的位置。因此，该方法可以帮助非制冷型红外相机更好地完成气体泄漏检测任务。更多还原

【Abstract】 Uncooled infrared cameras are widely used in the field of gas leak detection due to the advantages of low cost, long life and stable performance. An excellent image denoising algorithm can effectively improve the sensitivity and accuracy of detection. Combining deep learning and transfer learning techniques, an infrared image denoising method for gas leakage based on deep transfer learning is proposed in this work. Firstly, the convolutional neural network model is trained using a static scene dataset. Then some model parameters are fixed, and the model is retrained through simulating the gas dataset. Finally, a model suitable for denoising infrared images of gas leakage is obtained. The experimental results show that the method can denoise gas infrared images captured by uncooled infrared camera. The denoised images have obvious gas profile information, and the location of the leak source can be distinguished at the same time. Therefore, it is believed that the proposed infrared image denoising method can benefit uncooled infrared cameras better accomplish the task of gas leak detection.更多还原