Comparative Analysis Of Optimization For CNN-Based Feature Extraction And Bilstm Classification In Suspicious Video Behavior Recognition
DOI:
https://doi.org/10.29284/z7tpz648Keywords:
SGD, Adam, RMSprop, AdamW, CNN feature extraction, BiLSTM/LSTM.Abstract
For suspicious human behavior recognition, temporal sequence modeling and efficient spatial feature extraction are essential component for recognition. For recognition excelled network, Convolutional Neural Networks (CNNs) is used for extracting the hierarchical spatial features whereas Bidirectional LSTM (BiLSTM) networks are used for temporal dependencies thus training convergence, generalization, and computational efficiency are all greatly impacted by the optimization selection algorithms. Optimization algorithms like Adam, RMSprop, AdamW, Stochastic Gradient Descent (SGD), Nadam and Adadelta types are used in this research for comprehensive comparative analysis and then the optimized extracted features are given to CNN-LSTM/BiLSTM architecture for suspicious behavior classification. Each optimizer is evaluated carefully across key areas, such as how quickly converges, its accuracy, ability to generalize, its training stability, hyperparameter sensitivity, and overall computational cost. The evaluated optimized features are tested on three benchmark datasets viz: Violence Detection dataset, UT-Interaction dataset, and Custom Action dataset by adaptive learning rate approaches significantly outperformed the classical optimization methods. This optimization methods offers thorough convergence analysis, mathematical formulations, practical deployment guidelines, and an open-source implementation that facilitate research. This research offers solid, evidence-based suggestions for selecting the right optimizer in suspicious human behavior, which can cut down the training time and boost the accuracy.
References
[1]. M. Tan and Q. V. Le, “EfficientNet: Rethinking model scaling for convolutional neural networks,” Proc. ICML, 2019.
[2]. C. Szegedy et al., “Inception-v4, Inception-ResNet, and the impact of residual connections,” IEEE Conf. CVPR, 2017
[3]. S. Kornblith, J. Shlens, and Q. Le, “Do better ImageNet models transfer better?” Proc. CVPR, 2019.
[4]. Q. Dou et al., “Federated deep learning for medical imaging,” Nature Machine Intelligence, 2019.
[5]. J. Carreira et al., “A short note about the Kinetics-700 dataset,” arXiv 2019.
[6]. D. Tran et al., “A closer look at spatiotemporal convolutions for action recognition,” Proc. CVPR, 2018.
[7]. J. Zhao et al., “Multi-level spatiotemporal representation learning for video action recognition,” IEEE Trans. Image Process., 2020.
[8]. A. Piergiovanni and M. Ryoo, “Representation flow for action recognition,” Proc. CVPR, 2019.
[9]. E. Sterling and A. Gershon, “Attention-augmented recurrent networks for video temporal modeling,” IEEE Access, 2020.
[10]. S. Li et al., “Bi-directional recurrent networks for human action classification,” Pattern Recognition Letters, 2019.
[11]. F. Yang et al., “Attention-based LSTM for human action recognition,” Neurocomputing, 2019.
[12]. Y. Zhang et al., “Hybrid CNN–RNN architectures for efficient video classification,” IEEE Trans. Ind. Informat., 2020.
[13]. Y. Liu et al., “Global–local temporal modeling for action recognition,” IEEE Trans. Circuits Syst. Video Technol., 2021.
[14]. X. Li, “Hierarchical temporal representation networks for video understanding,” Proc. ECCV, 2020.
[15]. J. Lin, C. Gan, and S. Han, “TSM: Temporal Shift Module for efficient video understanding,” Proc. ICCV, 2019.
[16]. A. Bertasius, H. Wang, and L. Torresani, “Is Space-Time Attention All You Need for Video Understanding?” Proc. ICML, 2021.
[17]. M. Arnab et al., “Vivit: A video vision transformer,” Proc. ICCV, 2021.
[18]. A. Kumar et al., “Feature-boosted video transformers for action recognition,” IEEE Trans. Pattern Anal. Mach. Intell., 2023.
[19]. I. Loshchilov and F. Hutter, “Decoupled weight decay regularization (AdamW),” Proc. ICLR, 2019.
[20]. J. Luo et al., “Adaptive gradient methods with dynamic bound of learning rate,” Proc. ICLR, 2019.
[21]. H. Shin et al., “On the convergence of adaptive gradient methods,” IEEE Trans. Neural Netw. Learn. Syst., 2021.
[22]. C. Wilson et al., “The marginal value of adaptive gradient methods,” Proc. NeurIPS, 2017.
[23]. A. Chatterjee et al., “On the generalization performance of adaptive methods,” Proc. ICML, 2020.
[24]. J. Zhang et al., “Are adaptive optimizers beneficial for training deep neural networks?” Proc. AAAI, 2020.
[25]. D. Zhou et al., “Understanding generalization of adaptive gradient methods,” Proc. NeurIPS, 2020.
[26]. M. Sultani, Z. Chen, and M. Shah, “Real-world anomaly detection in surveillance videos,” Proc. CVPR, 2018 (foundational dataset; keep if needed).
[27]. S. Ramachandra and M. Jones, “Street scene: A new dataset and evaluation protocol,” Proc. WACV, 2020.
[28]. Z. Wu et al., “Learning to anticipate abnormal events via object-centric prediction,” Proc. CVPR, 2020.
[29]. P. Singh, A. Majumdar, and A. Nigam, “Human action recognition using CNN-RNN models: A comparative analysis,” IEEE Access, 2022.
Downloads
Published
Issue
Section
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
This work is licensed under a Creative Commons Attribution 4.0 International License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
