Novel Transfer Learning Attitude for Automatic Video Captioning Using Deep Learning Models

Abstract

Objectives: To generate the captions for the videos with less time complexity and high accuracy and also to create captions for each input video frame with particular timestamps. It will be utilized in the crime branch and hearingimpaired people will learn about the happenings of the video fruitfully. Methods: The proposed approach experiments with Transfer learning techniques. Modified Inception v3 and Resnet 50 networks are designed to compare the results. The standard MSVD Dataset is utilized to demonstrate the architectures. The performances are compared with the standard performance metrics. Findings: The inception v3 model works better than the Resnet 50 architecture for video captioning tasks. It provides the best accuracy at 99.83% with captions for the given input videos than Resnet 50 model. The MSVD dataset is more suitable for the demonstration of the video captioning task. Novelty: The two proposed models are modified based on the working of the video captioning tasks. The aggregation of some layers boosts the performance of the models more than ordinary models. Keywords: Artificial Intelligence; Automatic Captioning; Transfer Learning; Frames; Inception V3; Residual Network50 Model