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

This study addresses the pressing need for high efficiency and low resource consumption in crime pattern recognition within public safety governance by proposing a lightweight deep learning model known as the lightweight crime recognition network (LCRNet). Designed to provide intelligent support for forecasting and case classification, LCRNet integrates a Transformer encoder and convolutional neural network. To optimize performance, the model introduces simulated annealing sparsity (SAS) into the multi-head self-attention of the Transformer architecture, thus effectively reducing computational overhead while maintaining accuracy. Experimental results indicate that LCRNet achieves an accuracy of 97.76% on real-world crime data from Los Angeles and demonstrates strong generalizability in cross-dataset testing. Additionally, ablation studies and visualizations of the sparsity process confirm the effectiveness of SAS. This research provides a practical solution for efficient crime pattern recognition and edge device deployment in public safety, and our future work will focus on enhancing model interpretability and validating adaptability in resource-constrained environments.