Modified Hybrid Pooling on FaceNet Embedding to Enhance a Surveillance-Based Suspect Detection System

Abstrak

Criminality is increasing daily. CCTV cameras are strategically installed to monitor and regulate unlawful activity through efficient video surveillance. Video analysis by security personnel and conventional image processing are limited, particularly when individuals wear facial coverings such as glasses or masks. FaceNet is a face recognition system that is constructed on CNN, which often has difficulties in generalizing from unseen data. Adopting specialized pooling methods, such as hybrid pooling, can enhance the generalization of CNNs. This work aims to illustrate the use of hybrid pooling in FaceNet embeddings for suspect identification and to improve generalization in CNNs. Hybrid pooling is a probabilistic combination of max and average pooling. The proposed model outperformed the traditional pooling model (max and average pooling) across nearly all evaluation metrics when hybrid pooling substituted max pooling in the Reduction-B block of FaceNet embedding, achieving an alpha value of 0.5, an accuracy of 85.96%, an F1-score of 84.56%, a recall of 85.16%, and an AUC score of 90%. The experiment’s outcomes indicate that the alpha value and the choice of hybrid pooling inside a certain block significantly influence the model’s performance. The proposed method has improved the generalizability of FaceNet embeddings for suspect identification. Future research may investigate larger datasets, including scenarios that generate them for outdoor environments. Moreover, this system can be applied in real-world scenarios, such as supporting law enforcement in accelerating suspect identification in public areas, thereby supporting more efficient prevention and investigation.

Identifier

• OpenAlex: https://openalex.org/W7163745963
• DOI: https://doi.org/10.62527/joiv.10.3.4143