Abstract:
"Inexperienced YouTube creators struggle to choose the right video thumbnails, especially in areas with a dearth of qualified specialists. Effective thumbnail selection tools are needed to help content creators overcome these obstacles, but current tools focus primarily on producing new thumbnails and lack capabilities for assessing already-created thumbnails.
To address this gap, the authors propose a deep learning method for thumbnail rating. The method is trained using a hybrid ensemble deep learning architecture that uses ResNet50 and InceptionV3 as sub models embedded inside a neural network. The experiments were conducted using various augmentation techniques, architecture layer changes, and hyperparameter tuning.
The proposed model achieved an initial precision score of 0.79 and a final recall score of 0.82, resulting in a marginally higher overall improvement. The authors also compared the proposed model to a number of currently used approaches and found that it outperforms them in terms of overall performance and generalizability.
The authors conclude that the proposed deep learning method is a promising approach for thumbnail rating, especially in settings where annotated data is scarce. The method can be used to help content creators select better thumbnails for their videos, which can lead to increased views and engagement."
