Recaptured Image Forensics Using Transfer Learning

Recaptured Image detection is a field of security forensics that deals with detecting the originally captured image from its reimaged counterpart. The different algorithms that have been proposed to differentiate the original and recaptured images work with great accuracy on a large dataset and high-definition images. But the challenge remains for small-scale and small- size images as the existing algorithms required customized hand engineering whilst dealing with different datasets. Our proposed system is a dual learning model, which takes into account the enormous research for computer vision and applies concept of transfer learning for recaptured images detection. The ensemble of 6 pre-trained networks, namely AlexNet, Res Net, InceptionNet, GoogleNet, SqueezeNet, and MobileNet, has been compared on same dataset. It will allow institutional level detection of image spoofing and the recaptured image forgery for small-scale images with considerable accuracy without requiring specialized preprocessing on their end. We have achieved the highest accuracy of 90 percent through our model containing Transfer Learned MobileNet and SVM classifier. As per the latest resources, this is the first instance of applying transfer learning for recaptured image detection.