Learning optimal deep prototypes for video retrieval systems with hybrid SVM-softmax layer

The research focuses on optimizing training time for video retrieval by producing optimized prototypes for a hybrid SVM-softmax regression classifier. A modified particle swarm optimization-based feature selection technique is used to reduce feature dimensionality. A multi-class SVM variant is used for initial weight assignment, and each query video’s classified value is compared with the prototype features. The optimized prototype learned hybrid classifier improves top 1 retrieval accuracy in L2 distance on benchmark datasets, surpassing previous state-of-the-art methodologies.