| The purpose of person re-identification technology is to match person images with the same identity from different perspectives and different postures taken under different cameras,so as to realize cross-camera recognition of person.As an important research branch,cross-resolution person re-identification played a great role in public security management,tracking suspects,crime prevention and so on.Although great progress has been made in cross-resolution person re-identification,the following problems still exist in practical application scenarios: First,due to the difference in shooting distance,shooting angle and camera parameters,the captured person images had different resolutions.If different resolution images were matched directly,it will lead to a decrease in recognition accuracy.Secondly,the interference of complex background information will also lead to a decline in recognition accuracy.This paper combined image superresolution technology and person re-identification technology to study the above problems.The main research contents and innovations are as follows:First of all,to solve the problem that the resolution of person images obtained in actual scenes is not fixed,this paper proposed a cross-resolution person re-identification method based on meta-learning super-resolution network,which includes meta-learning super-resolution network and person re-identification network based on local feature fusion.In the super-resolution network,the idea of meta-learning was used to construct a meta-learning upsampling module to achieve multiple scale super-resolution of images.In the person re-identification network,extracted local features of an image and introduced an adaptive local feature fusion module.This structure strengthened the expressive ability of features through multiple feature interactions.The experimental results showed that the algorithm can re-identify images with different resolutions and achieve high accuracy.Secondly,in order to reduce the accuracy caused by the interference of background information in cross-resolution person re-identification,the Meta SRRe ID was improved and a cross-resolution person re-identification method based on graph convolution network was proposed.In the super-resolution network,the multi-head self-attention mechanism of Transformer was introduced to capture the global dependency,which makes the network pay more attention to the important areas of the image and generate a super-resolution image which was more convenient for recognition.In the person reidentification network,the human key point detection model was introduced,which can effectively reduce the interference of complex background information while extracting the features of human key points.Then the key point features were regarded as the nodes in the graph convolution and formed the human body topology graph.Finally,the proposed feature fusion method based on graph convolution network was used to transfer and aggregate the information between adjacent nodes and fused into a new feature representation.The experimental results showed that the algorithm can effectively suppress the interference of complex background information,thus improving the accuracy of cross-resolution person re-identification.Finally,a cross-resolution person re-identification software platform based on Py Qt5 was designed and implemented,and the method proposed in this paper was encapsulated to realize cross-resolution person re-identification and intuitively display the recognition results. |
