Advertisement Detection in broadcasted videos using Transfer Learning and Support Vector Machine

Main Article Content

Namrata Dave, Dr. Mehfuza S. Holia


Deep learning using convolutional neural networks is emerged as the best approach for object detection. Performance of convolutional neural network depends on its architecture as well as large dataset for training. To train the model with small dataset and get most accurate results we have adopted transfer learning approach. We have used Alexnet model to train on dataset consisting of keyframes of news and advertisement frames obtained from news video of DD Girnar, ETV Gujarati, TV9 news, Sandesh channels. Using learned weights of well-defined network trained on very large dataset, proposed work achieved very good results in related tasks by adding new dataset in training. To achieve good results, experiments were performed on different models with different layers to generate final results. In this paper, experiments performed with pretrained Alexnet model for training as well as classification is presented with the results obtained. To achieve better accuracy without compromising training time, another transfer learning approach with Alexnet as feature extractor and SVM as a classifier is proposed. To classify images, Support Vector Machine along with Bayesian optimizer is used to improve classification performance. Proposed approach gives 99.2 percent accuracy on the dataset of news video.

Article Details