Infrared Image Target Detection Based On Domain Adaptation

The rapid development of the railway industry has brought about an increasingly dense railway network,and the safe operation of trains is also particularly important.In the process of running the train,foreign objects invade the perimeter of the railway and occasionally cause the train to be shut down and even cause huge accidents,which poses a great threat to people’s lives and property safety.Therefore,how to achieve rapid and effective detection of railway foreign target intrusion limits and prevent railway traffic accidents has become a research hotspot.Researchers have applied deep learning and computer vision theories to the detection of railway foreign target intrusion limits,and have achieved rich research results.However,most of the achievements in this field are obtained from RGB images during the day,and there are few researches on infrared mode images at night.The number of infrared image samples of railway scenes is small,and infrared samples of abnormal targets are more difficult to obtain,and deep learning has a strong dependence on the number of samples.Therefore,conventional deep learning target detection methods cannot meet the requirements of railway intrusion detection in night infrared mode.In response to this problem,this paper is based on the deep learning target detection algorithm,fused with the idea of domain adaptation,and proposes an infrared image target detection algorithm based on domain adaptation,which solves the problem of lack of infrared image samples and improves the deep learning target detection.The foreign target detection accuracy of the model in the night railway scene.The method in this paper consists of two stages: First,based on the pixel-level domain adaptation method,this paper uses Cycle GAN(Cycle Consistent Generative Adversarial Networks)as the basic network,and proposes a C-Cycle GAN model.Infrared style transfer,a large number of infrared style image samples are generated,and data supplementation is completed;secondly,this paper combines the domain adaptation idea with the SSD(Single Shot Multibox Detector)algorithm,and proposes a depth based on the feature adaptation domain.Learning target detection methods,designing a loss function based on Maximum Mean Discrepancy(MMD),realizing the alignment of the source domain and target domain features,reducing the difference between domains,and improving the accuracy of target detection in infrared images of the target domain rate.In order to verify the effectiveness of this method,this paper conducted multiple experiments to compare the detection effects of the traditional SSD algorithm and the algorithm of this paper.Experimental results show that the detection accuracy of the infrared image target detection algorithm based on domain adaptation proposed in this paper can reach 90.4%,which is 3.9% higher than the traditional SSD algorithm.The problem of low accuracy of target detection in railway scenes in the night infrared image mode is better solved.