Lightweight deep learning model for crime pattern recognition based on transformer with simulated annealing sparsity...

Area of Science:

• Artificial Intelligence
• Computer Science
• Public Safety Technology

Background:

• Public safety governance requires efficient crime pattern recognition.
• Existing models often have high resource consumption, limiting deployment on edge devices.
• Intelligent forecasting and case classification are crucial for effective crime management.

Purpose of the Study:

• To propose a lightweight deep learning model, LCRNet, for efficient crime pattern recognition.
• To reduce computational overhead and resource consumption in crime analysis.
• To provide intelligent support for crime forecasting and classification.

Main Methods:

• Integration of a Transformer encoder and a convolutional neural network (CNN).
• Introduction of simulated annealing sparsity (SAS) into the Transformer's multi-head self-attention mechanism.
• Optimization of model performance for accuracy and efficiency.

Main Results:

• LCRNet achieved 97.76% accuracy on real-world Los Angeles crime data.
• Demonstrated strong generalizability across different datasets.
• Ablation studies and visualizations confirmed the effectiveness of SAS in reducing computational overhead.

Conclusions:

• LCRNet offers a practical and efficient solution for crime pattern recognition.
• The model's lightweight design facilitates deployment on edge devices in resource-constrained public safety environments.
• Future work will focus on enhancing model interpretability and adaptability.