"PsoGhost-crime-classification-dataset"

"Neural network architecture search (NAS) is considered a robust methodology for exploring high-performance network architectures. However, current methodologies exhibit persistent limitations, including a single form of network architecture design or limitations in the search space. To solve these problems, this study proposes an efficient network architecture search algorithm based on deep particle coding. Firstly, the improved Ghost block is used as the basic unit of architecture search, and the structured redundancy within the network is exploited through a dynamic attention selection mechanism to jointly optimize computational efficiency, model complexity, and representational capability in the search space. In addition, we develop a deep wrapped particle coding strategy with an outside-in hierarchical search mechanism to enhance architecture exploration within the baseline framework. The experimental results on benchmark datasets and the self-collected Crime dataset demonstrate that the proposed framework can automatically construct classification models with low complexity and strong discriminative performance, suggesting that structured redundancy may provide useful inductive bias for lightweight architecture search. The code is available at https:\/\/github.com\/wangdianwei\/Pso-GhostNet. "