Fig 1.
Residual behavior.
Fig 2.
Inverted residual bottleneck block.
Fig 3.
Proposed architecture for human action recognition on selected action datasets.
Fig 4.
Sample classes from the HMDB51 dataset.
Fig 5.
Sample classes from UCF101 dataset [40].
Fig 6.
Proposed inverted residual parallel block.
Fig 7.
Self-attention module for the features learning.
Fig 8.
Complete architecture of proposed inverted bottleneck residual with self-attention CNN.
Fig 9.
Detailed, layered architecture of proposed Inverted Bottleneck Residual with Self-Attention (InBRwSA) for action recognition.
Fig 10.
Proposed architecture testing phase.
Table 1.
Proposed architecture classification accuracy using the HMDB51 dataset.
Fig 11.
Confusion matrix for SWNN classifier using proposed architecture.
Table 2.
Proposed architecture classification accuracy using UCF101 Dataset.
Fig 12.
Confusion matrix for SWNN classifier using proposed architecture using UCF101 dataset.
Fig 13.
Ablation study 1- pre-trained deep learning-based comparison for the selected datasets.
Fig 14.
Ablation Study 2- Confidence interval-based analysis of proposed architecture on selected datasets.
Table 3.
Comparison with existing SOTA techniques.
Fig 15.
Proposed architecture prediction results using video sequences.