Video Content Analysis to Distinguish Between Violence and Nonviolence Using ResNet101v2 and LSTM

Video content analysis is the most important field in which video classification techniques provide a good result to differentiate ‘Violence’ with ‘Nonviolence. Although essential, the traditional ResNet50 model is not as successful in detecting this series compared to VGG19 and InceptionV3 with a validation accuracy of 83.87%. To alleviate this problem, LSTM networks were introduced to make video data more susceptible for sequential processing and trained with the same spatial CNN vanish network reaching an accuracy of 97.54%, but a validation accuracy of slightly lower: only 80.65%. This indicates that some over fitting also occurred in this approach as well which suggests dubious reliability on our LSTMs; A new hybrid method to utilize the spatial natural of image using ResNet101v2 and temporal sensing nature with LSTM was suggested in this paper. This hybrid achieved 99.18% training accuracy and a validation of only 77.42%. The difference highlights the difficulty of training data that generalizes well (to new, unseen examples) and improves performance) did Receive Memory Warning () the results highlight the continued hurdle in computational video analysis: designing models that balance high predictive accuracy with their ability to generalize across disparate datasets.