A Light Weight Depthwise Separable Layer Optimized CNN Architecture for Object-Based Forgery Detection in Surveillance Videos.

This article focuses on developing a passive approach for detecting spatio-temporal object-based forgery, specifically object removal, in surveillance videos. It introduces a novel max averaging temporal windowing technique to extract motion residue frames that enhance the localization of forged regions. These motion residues are then analyzed using a lightweight, optimized depth-separable layer convolutional neural network (DS-layer CNN), which achieves high accuracy (98.60% at frame level and 99.01% at video level) with fewer trainable parameters and faster training compared to existing models. The method is validated on the SYSU-OBJFORG dataset for object removal forgery and further tested on the publicly available REWIND dataset for copy-move forgery, demonstrating its potential generalizability. The study emphasizes the model’s computational efficiency and suitability for real-time applications, with future work aimed at improving forgery localization and broader generalization.