Optimized deep learning model for spatio-temporal detection and localization of object removal video forgery with multiple feature extraction
Lakshmi Kumari, PRASAD K.V · Research Square · 2022
Abstract Video forgery (VF) is an approach for manipulating the fake videos by modifying, coordinating or generating new contents among the video sequence. The identification of this type of video forgery is complex. For extracting multiple features, developing a novel approach still remains major challenge in this area. In this work, an optimized deep learning (DL) model for the video forgery detection (VFD) and localization with multiple feature extraction is proposed. Initially, key-frame extraction takes place with the aid of Gaussian mixture model (GMM) to extract frames from the forged videos. Then, pre-processing stage is manipulated for the conversion of RGB frame into grayscale image. To study the nature of the forged videos, there is a need of extracting multi-features from the pre-processed frames. In our proposed study, SURF, PCA-HOG, MBFDF, COA and PRG features are extracted. The dataset used for proposed work is collected from REWIND of about 80 forged and authenticated videos. With the help of DL approach, video forgery can be detected and localized. Thus, this research mainly focus on the detection and localization of forged video based on ResNet152V2 model hybrid with Bi-GRU to attain the maximum accuracy and efficiency. The performance of this model is finally compared with existing approachesin terms of accuracy, precision, F-measure, sensitivity, specificity, FNR, FDR, FPR, MCC and NPV. The proposed methodology assures the performance of96.17% accuracy, 96% precision, 96.14% F-measure, 96.58% sensitivity, 96.5% specificity, 0.034 FNR, 0.04 FDR, 0.034 FPR, 0.92 MCC, 96% NPV respectively. Along with is, the mean square error (MSE) and peak-to-signal-noise ratio (PSNR) for GMM model attained about 104 and 27.95 respectively.