Target Detection Of Infrared Small Sample

With the popularization of computer and the progress of digital image processing technology,computer vision technology has developed rapidly.Compared with traditional methods of target detection and recognition,the method of deep learning has realized automation,intelligence and high efficiency,and has higher detection and recognition accuracy.Compared with visible image,infrared image has stronger anti-interference ability,and target detection of infrared images has become more important,especially in military fields and harsh environments.Nowadays,target recognition and detection technology mainly concentrates on the field of visible image.The principle of visible image is different from that of infrared image.Therefore,the accuracy of infrared image detection and recognition will be reduced if the depth neural network trained by visible image is directly used.However,the training of a new network by using infrared image directly will face less samples of infrared image.Aiming at the problems mentioned above,this paper proposes an infrared-energy-to-color(IETC)method,a novel infrared-to-visible framework based on Cycle-Consistent Generative Adversarial Networks(CycleGAN)to transform infrared images to visible images,and uses the target detection network trained totally by visible images to test generated images.IETC can extract the infrared energy information and learn the correlation of energy and color feature to generate the visible images fusing energy information.IETC includes a detail-enhanced pre-processing module and an energy information extraction structure.The pre-processing module uses the guided filtering algorithm to process the infrared image,enhances the image contrast,enriches the image details,and better preserves the image edge.The energy information extraction structure is constructed by introducing an energy bridge to connect the infrared energy to the generated visible image and improve the structure of CycleGAN to make the network model suitable for the conversion.The results show that the proposed model can generate a close-to-reality visible image with stereoscopic target and background.Compared with the infrared image directly detected by the target detection network trained by the visible image,the visible image generated by the method proposed in this paper can obtain more accurate detection results.