This paper provides a comprehensive analysis of several neural network architectures designed for the detection of generative image inpainting. The central challenge addressed is the increasing sophistication of inpainting technologies, driven by generative AI, which renders traditional forensic methods based on simple statistical or visual artifacts obsolete. This paper investigates a feature-centric hypothesis: that visually seamless inpainting manipulations introduce consistent, subtle, and detectable artifacts in specific high-frequency and statistical domains. The primary domain of interest is the feature space defined by the Dual-Tree Complex Wavelet Transform (DTCWT).
The core of this report details the methodology and evaluation of six novel deep learning architectures, each custom-designed to operate on these wavelet-domain features. These architectures explore diverse design philosophies, including dense feature aggregation (Custom UNet++), explicit global context integration (UNet++ with Global Average), recurrent spatial modeling (ConvLSTM), unsupervised generative anomaly detection (Variational Autoencoder), and channel-specific processing (StackDeepAll).
A rigorous experimental evaluation was conducted on a custom inpainting detection dataset. The quantitative results demonstrate the clear superiority of the StackDeepAll architecture. This model, which processes each of the 12 DTCWT channels through independent, parallel UNet-like encoders before fusion, achieved a state-of-the-art Intersection over Union (IoU) of 0.8126 in validation. This result significantly surpassed all other proposed models, including a stable Variational Autoencoder (0.3711 IoU) and overfitting-prone ConvLSTM models (0.4514 IoU). The findings establish that for high-dimensional, engineered features like complex wavelets, a channel-specific processing paradigm that avoids premature feature-blending is a powerful and highly effective design strategy. This feature-centric approach is validated as a potent alternative to contemporary end-to-end spatial-domain models.
