Research On Key Technologies Of Infrared-Visible Light Cross-Modality Face Recognition

In the dark scene at night,the face images captured by the ordinary monitoring camera are generally poor in definition and very dark,while the near-infrared(NIR)camera can capture the NIR face with high definition at night.Therefore,NIR imaging technology has broad application prospects in the monitoring system.However,the pre-registered face in many applications is the face under visible light(VIS),which needs to match the NIR face captured by the camera with the VIS face in the photo library,so as to determine the identity of the NIR face image,which is of great significance to maintain social security and assist the police in solving cases.In order to make better use of the advantages of infrared imaging,the thermal infrared camera is used to detect whether someone passes or approaches at night at a long distance,which is conducive to early warning;NIR camera is used to shoot clear faces and identify them in close range,which helps to lock suspects.In this thesis,the super-resolution reconstruction of thermal infrared images and the cross-modality face recognition of NIR-VIS are studied for the applications of infrared imaging.The specific research contents are as follows :(1)In the process of imaging or transmission,the resolution of thermal infrared image is low due to many factors such as environment,component performance and so on.In order to improve the resolution of thermal infrared images,this thesis proposes a superresolution reconstruction method of thermal infrared images based on channel attention and transfer learning.This method first designs a deep convolutional neural network,which integrates channel attention mechanism to enhance the learning ability of the network,and uses residual learning to accelerate the convergence of the network.Then considering the problem that high-quality thermal infrared images are difficult to collect and the number is insufficient,the network training is divided into two steps: The first step is to pre-train the network model with many natural images,and the second step is to fine-tune the pre-trained model with a small number of high-quality thermal infrared images by using the knowledge of transfer learning.Finally,multi-scale detail filtering is used to further improve the edge clarity and visual effect of the reconstructed thermal infrared image.The experimental results show that the proposed method has obvious effect on improving the resolution of thermal infrared images.(2)In order to improve the accuracy of NIR-VIS cross-modality face recognition,this thesis designs a NIR-VIS cross-modality face recognition method based on modality gap decomposition.Firstly,the lightness component(V component)of HSV space of VIS image is taken as an auxiliary modality to reduce the difficulty of network learning.The lightness component retains the structural information of the VIS image and is similar to the color information of the NIR modality,so the huge gap between the NIR modality and the VIS modality of network learning is decomposed into two smaller gaps in learning.Secondly,we input the data of three modalities into the weight sharing network and train them under the joint guidance of the cross-modality gap segmentation loss and the intramodality gap loss designed in this paper.The experimental results show that the proposed method has obvious effect on improving the accuracy of NIR-VIS cross-modality face recognition.