基于GAN的红外小目标检测合成数据增强技术研究

Research on GAN-Based Synthetic Data Augmentation Technology for Infrared Small Target Detection

近年来,卷积神经网络在红外小目标检测方面取得了良好的效果;但由于该技术中公共训练数据数量有限,在一定程度上导致分析结果不够精确;为了解决训练数据的稀缺问题,提出了一种生成红外小目标检测合成训练数据的方法;采用生成对抗网络框架,其中合成背景图像和红外小目标在两个独立的过程中生成;在第一阶段,通过将可见光图像转化为红外图像来合成红外图像;在第二阶段,将转换后的图像上植入目标掩码,将所提出的强度调制网络合成了真实的目标对象,可以从进一步的图像处理产生;实验结果表明,当使用由真实图像和合成图像组成的数据集训练各种检测网络时,检测网络比只使用真实数据产生更好的性能,优化了传统数据处理方法,使得训练结果更为精确。

Recently,convolutional neural networks have made a great progress in infrared small target detection.However,due to limited public training data in this technology,the analysis results are inaccurate a certain extent.To address the scarcity of training data,this paper proposes a method for generating synthetic training data for infrared small target detection,and adopts a generative adversarial network framework,where the synthesized background image and infrared small targets are generated in two independent processes.In the first stage,the infrared images are synthesized by converting visible light images into infrared images.In the second stage,the target mask is implanted on the converted image,and the proposed intensity modulation network is synthesized into the real target object,which can be generated from further image processing.Experimental results show that,when the dataset on real and synthetic images trains various detection networks,the detection network of the real and synthetic images has better performance than that of only real images,optimizing the traditional data processing methods and making the training results more accurate.