Self-Supervised CNN for Video Double Compression Detection via Artifact Prediction

We propose a self-supervised learning framework for identifying secondary compression artifacts in H.264 videos, which tackles the challenge of scarce labeled data in forensic analysis. The proposed method includes a Compression Artifact Prediction Module (CAPM) that trains a convolutional neural network (CNN) beforehand to estimate synthetic compression parameters from unlabeled video frames, which helps the model acquire discriminative features without requiring extensive annotations. The system produces synthetic artifacts by modeling controlled re-compression scenarios with different quantization parameters and frame rates, which act as self-supervision signals for pre-training. The CAPM employs a multi-task loss to jointly estimate compression parameters, frame type distributions, and motion vector magnitudes, thereby capturing intrinsic patterns of compression artifacts. Additionally, a module dedicated to temporal consistency combines spatiotemporal features from multiple frames by applying 3D convolutions, whereas a contrastive detection head improves the distinction between primary and secondary compression traces. The last detection phase merges transformer-driven self-attention and a binary classifier to capture extended dependencies and strengthen decision reliability. Experiments show that our method attains good performance in secondary compression detection while markedly decreasing dependence on annotated data.