Deep Learning Based Cross-Modality Recognition Algorithm

Biometric recognition technology is a technique that uses human biometric characteristics for identity authentication and recognition,and has been widely applied in various real-word scenarios.Conventional biometric recognition techniques,such as face recognition and person re-identification,mainly focus on matching problems between visible(rgb)images.These methods can achieve very high accuracy when the environmental conditions are good,but degrade dramatically under low-light conditions.With the development of surveillance equipment and the increasing demand of security,some cameras can switch between visible mode and infrared mode under strong light and weak light conditions respectively,making it possible to capture high-quality images in dark indoors or at night.Therefore,studying cross-modality biometric recognition techniques that can match visible images and infrared images has important theoretical significance and practical value.This topic divides crossmodality biometric recognition into two subtasks for research:cross-modality person re-identification and cross-modality face recognition.In addition to the problems in traditional person re-identification,such as occlusion,background clutter,pose,etc.,cross-modality person re-identification encounters huge modality variations.This paper proposes a method of adaptive part capture and global-local feature fusion.This method can automatically locates the foreground region of pedestrians and extracts local features by using attention mechanism,and fuses local features with global feature to obtain pedestrian features with strong discrimination and robustness.Moreover,this paper adopts cross-modality contrastive learning loss function to make the model modality-invariant,thereby enhancing features.Experimental results show that our method can achieve better results in more common data distribution scenarios.Cross-modality face recognition also faces the problem of large differences in data distribution between different modalities.Currently,there is a lack of a general and effective baseline model in this field.This paper draws on some training tricks from cross-modality person re-identification,and constructs a relatively simple and powerful baseline model.Furthermore,this paper proposes a hierarchical identity-modality disentanglement method,which decouples and purifies the features output by the network blocks step by step,so that the network can output features containing only identity-related information.To enhance the capability of feature disentanglement,this paper utilizes an identity-modality disentanglement loss function to make the identity-related feature distribution more compact and make the modality-related feature distribution more dispersed.Experimental results demonstrate that our method can effectively improve the performance of the model on cross-modality face recognition tasks.