Research On Visible-Infrared Cross Modality For Person Re-Identification Method

In recent years,visible-infrared modality Person Re-identification(Re-ID)has been widely applied to video surveillance,human-machine interconnection,intelligent security and other fields.However,due to the difference in visual appearance between images and the defects of infrared images such as low resolution,so that the study of cross modality person Re-ID is in the infant stage at present.Through deeply analysis of the cross modality person Re-ID problems,two solutions are proposed,which are feature pyramid random fusion network and multi-modality learning of hybrid loss function.It effectively solves problems of visible-infrared person Re-ID.Extensive experiments demonstrate the effectiveness of our approach.Main researches of our work is as follows:1.A feature pyramid random fusion structure is proposed for the limitation of single scale feature learning and the complexity of multi-modality feature fusion.Firstly,we introduce a super-resolution reconstruction method to handle training disturbance caused by visual blur from infrared images.Secondly,our work implements high-resolution with multi-semantic information as a single output by mixing strong semantic features at toplevel to strong geometric detail features at bottom-level.Finally,integrating the advantages of learning global and local,a random fusion mechanism of convolutional layers from multi-modality and multi-hierarchy is designed.This method completes the task of image multi-scale semantic learning,and the number of parameters involved in the operation is not high,thus effectively solves the problems of cross-scale multimodality learning of Re-ID.Experimental results testify that the accuracy of our work is better than existing methods,and the model is smaller in size and shorter in training period.2.The vital factor affecting the performance of cross-domain Re-ID is gaps within intra-modality visual appearance and inter-modality heterogeneity issues.Thus we propose a hybrid loss function of classification learning model.Firstly,the SVD layer is designed based on the characteristic of the singular matrix.Secondly,the cross-entropy is determined as the basic loss function,then a bilinear interpolation skill is used to explicit the gap.Finally,we utilize the bias from the three modalities of RGB-RGB,IRIR,and RGB-IR to modify the network.The experimental results testify that our algorithm is effectively correcting the model.