Image Authenticity Detection via YOLOv5 with Frequency Domain Feature Fusion

Abstract

With continuously accelerated developments in digital image forgery technologies, authenticity verification in images is a key research direction for computer vision. The proliferation of forged images poses severe threats to information security, media credibility, and social stability. Current detection methods face limitations in real-time performance and generalization, struggling against increasingly sophisticated forgery techniques. This study proposes an efficient deep image authenticity detection framework based on the YOLO object detection model, aiming to achieve high-precision, real-time forgery region localization and classification. Traditional image authenticity detection methods, including manual inspection, watermark detection, and physical inconsistency analysis, suffer from poor generalization and low efficiency. To address these limitations, we propose an automated deep image authenticity detection algorithm based on the YOLO model. Using the IMDb-Wiki dataset from ETH Zurich's Computer Vision Lab, our method involves: (1) constructing a YOLOv5 model for portrait region identification, accurately extracting five human body parts (head, hand, body, leg, foot); (2) further localizing and cropping facial regions, enhancing image features through five preprocessing techniques for feature extraction: original image retention, Fourier transform, combined Fourier and high-pass filtering, Gaussian blur, and residual map analysis; (3) analyzing image frequency domain features using Convolutional Neural Networks (CNN) for comprehensive authenticity determination. Our experimental results demonstrate that the proposed framework achieves superior performance in terms of accuracy and real-time capability compared to existing methods. The integration of YOLOv5's multi-scale feature extraction with frequency domain analysis effectively captures subtle forgery traces, providing a robust solution for digital forensics applications.