Transfer-Learning and YOLO V7 Hybridised for Human Cropping for Deepfake Detection Algorithms

The advent of deep learning technology has created images and videos convincingly fake to appeal viewers enough in appearing close to real. Such image and video modalities, known as Deepfakes, can cause harm to individuals, communities, and countries. Detecting Deepfakes is essential to prevent their negative impacts, and several detection techniques that utilize deep learning algorithms have been developed over the recent past years. The algorithms reduce the target modality to dimensional latent space to include features important for subsequent reconstruction of the same modality from the developed latent distribution data. However, manual data preprocessing such as splitting frames and cropping regions from Deepfake videos for the facial features can be tedious and time-consuming. This article proposes YOLO V7 hybridised with transfer-learning to automate the cropping process in Deepfake videos and frames splitting, making it easier to extract facial features from videos. Various popular datasets have been used for Deepfake detection, and this article explores the benefits of the proposed approach. The results present a comparison of the proposed approach with state-of-the-art models from the levels of Precision, Recall, and mAP@.5 to demonstrate the robustness of the proposed technique. Notably, it is evident that our approach outperforms other models, boasting a precision value of 0.9009, a recall value of 0.8979, and an mAP@.5 value of 0.8745.