Research On Identity Recognition Based On Adaptive Feature Fusion

Communities are the building blocks of society and identification technology assumes a stabilizing role,effectively identifying potential problems that endanger community safety.In this paper,through the fusion of face and pedestrian feature information,we exploit the effective identity information after the interaction of multiple feature information in realistic and complex scenes and help the community to achieve accurate identification of the target identity.The following elements were mainly studied:1.Construction of a target recognition dataset.In this paper,surveillance data of different pedestrians under different cameras at different periods in different locations are collected to construct a target identification dataset.The constructed dataset includes a database of pedestrians in frontal,side,and head-down poses,as well as a database of faces corresponding to each pedestrian database.Analogous to the production scale of the public dataset Market1501,a pedestrian and face database with 1392 IDs was finally produced,with 696 IDs in both the training and test sets.2.A target recognition method based on gated feature fusion with central loss is proposed to address the problems of low accuracy of single feature recognition and non-compact features of target classification networks.First,using the gating mechanism,the network is instructed to evaluate the input face features and pedestrian features in terms of recognition contribution,set the score weights by the level of contribution,and then combine them to produce more robust identity features,adaptively solving the problem of decreasing recognition accuracy due to missing and blurred feature information.Secondly,a fusion loss function combining gated classification and central distance loss is proposed to guide the network to reduce the intra-class distance between features,increase the inter-class distance and train features that are more discriminative.Finally,validation on a self-constructed dataset shows that the complementarity between the strengths and weaknesses of multiple features can indeed cope with identity tasks in more realistic and complex scenarios.3.To address the problems of difficult medium-term feature fusion and insufficient useful information on time-domain features.An identity recognition method based on multi-feature mid-term fusion with frequency domain attention is proposed.Firstly,a multi-feature medium-term fusion network is designed to achieve the fusion of features from shallow to deep layers;Secondly,a feature scheduling module is proposed that uses channel attention to enable interaction between features,satisfies the conditio n that different network layers can be arbitrarily inserted between models without random initialization of training weights,and facilitates the slow fusion of multi-feature information.Finally,the discrete cosine transform is used to extend the face and pedestrian feature information in the time domain to the frequency domain for analysis.Based on the research basis that the traditional global average pooling is the lowest frequency component in the frequency domain,the high-frequency information is appropriately discarded and the low-frequency useful information is retained to provide more original feature information for the subsequent feature interaction work of channel attention and to achieve the full utilization of the target effective feature information.In summary,the two methods proposed and the dataset produced in this paper alleviate the challenge of insufficient accuracy of single-feature recognition at the level of multi-feature fusion,and provide technical and content supplements for identity recognition tasks in realistic scenarios,which are of some significance in promoting the development of research on identity recognition using face and pedestrian fea tures.