Learning clustered deep spatio-temporal prototypes using softmax regression for video information systems

For solving the data imbalance, dimensionality reduction, and optimizing the training time research problem, this research developed a method for producing an optimized number of prototypes using the clustering algorithms to provide training data to a logistic regression classifier for video retrieval. The optimization is based on the number of minimum class samples. This study uses clustering-based sampling to extract prototypes, which results in fewer training samples and shorter training times. The classified value for each query video is compared to the prototype feature’s classified value for each class in the database. According to the performance of the Clustered Prototype on pre-trained 3D Resnet features learned classifier, the top 1 accuracy of retrieval results concerning L2 distance (Euclidean distance) for both the datasets UCF101 and HMDB51 has improved above novel methodologies. The top 1 accuracy of the UCF101 dataset has been observed as 0.844 and the HMDB51 dataset as 0.5620. There is an improvement of 5.1% in the top 1 mean average precision of retrieval in the UCF101 dataset and 1.9% in the HMDB51 dataset than the original features without machine learning.