深度迁移学习预训练对红外尾流成像识别的影响

随着水下航行器噪声水平的不断降低，水下航行器形成的尾流红外成像特征就成为其主要可探测的特征源之一，利用水下航行器尾流的水面红外特征来探测水下航行器的踪迹逐渐发展成为一种新的探测方式。由于人工判别尾流特征的效率低，准确性不高，采用人工智能深度学习的方式能够得到较大的改善。本文以水下航行器尾流红外特征识别为研究核心，通过图像分类制作了混合类的样本集，利用迁移学习比较不同预训练网的对尾流的训练效果，讨论预训练网内部参数对尾流训练效果的影响，结合Faster-RCNN算法，最终测试对尾流的识别精度，在45个2类尾流的小样本集下，预训练之后的网络在识别准确度上增加了21.43%，误检率下降了2.14%，带有红外特征的图像在定位精准率上比可见光图像高18.18%。该预训练测试对未来研究尾流探测结合卷积神经网络的识别有一定的应用潜力。

The Influence of Deep Transfer Learning Pre-training on Infrared Wake Image Recognition

With lower underwater vehicle noise levels, the infrared imaging characteristics of underwater vehicle wake have become one of the main detectable sources. Using the infrared characteristics of underwater vehicle wakes to detect underwater vehicle traces has gradually developed into a popular detection method. Because of the low efficiency and inaccuracy of artificial wake characteristics identification, the adopted artificial intelligence deep learning method can be greatly improved. In this study, the infrared feature recognition of underwater vehicle wake is the primary focus. A sample set of mixed classes was made by image classification. The training effect of different pre-training networks was compared using migration learning. The influence of the internal parameters of the pre-training networks on the training effect of the wake was discussed. Finally, in the small sample set of 45 two kinds of wake, the recognition accuracy of the network after pre-training increased by 21.43%, the false detection rate decreased by 2.14%, and the positioning accuracy of the image with infrared characteristics was 18.18% higher than that of the visible image. This pre-training test has a certain application potential for future research on wake detection combined with convolution neural network recognition.