Enhancing Person Re-Identification through Attention-Driven Global Features and Angular Loss Optimization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Pedestrian re-identification (re-ID) faces challenges in accurately representing and discriminating pedestrian attributes.
• Existing methods struggle with feature representation and classification accuracy.

Purpose of the Study:

• To propose a novel method for person re-identification that improves feature representation and classification accuracy.
• To enhance the discriminability of pedestrian features for more robust recognition.

Main Methods:

• Integration of a Normalization-based Channel Attention Module with the ResNet50 backbone to prioritize key pedestrian features.
• Employment of dynamic activation functions to adaptively modulate ReLU parameters, boosting nonlinear expression.
• Incorporation of Arcface loss with cross-entropy loss for supervised training, promoting inter-class variance and intra-class coherence.

Main Results:

• Achieved a 1.28% increase in Rank-1 accuracy on the Market1501 dataset.
• Obtained a 1.4% increase in Rank-1 accuracy on the DukeMTMC-ReID dataset.
• Demonstrated improvements in mean average precision (mAP) by 1.93% and 1.84% on the respective datasets.

Conclusions:

• The proposed model effectively extracts robust pedestrian features, enhancing feature discriminability.
• The method leads to superior recognition accuracy in person re-identification tasks.
• The approach addresses limitations in pedestrian attribute representation and discrimination.