Generalizing to Out-of-Sample Degradations via Model Reprogramming

TL;DR

This work introduces the Out-of-Sample Restoration (OSR) task, which aims to develop restoration models capable of handling out-of-sample degradations, and proposes a model reprogramming framework, which translates out-of-sample degradations by quantum mechanic and wave functions.

Abstract

Existing image restoration models are typically designed for specific tasks and struggle to generalize to out-of-sample degradations not encountered during training. While zero-shot methods can address this limitation by fine-tuning model parameters on testing samples, their effectiveness relies on predefined natural priors and physical models of specific degradations. Nevertheless, determining out-of-sample degradations faced in real-world scenarios is always impractical. As a result, it is more desirable to train restoration models with inherent generalization ability. To this end, this work introduces the Out-of-Sample Restoration (OSR) task, which aims to develop restoration models capable of handling out-of-sample degradations. An intuitive solution involves pre-translating out-of-sample degradations to known degradations of restoration models. However, directly translating them in the image space could lead to complex image translation issues. To address this issue, we propose a model reprogramming framework, which translates out-of-sample degradations by quantum mechanic and wave functions. Specifically, input images are decoupled as wave functions of amplitude and phase terms. The translation of out-of-sample degradation is performed by adapting the phase term. Meanwhile, the image content is maintained and enhanced in the amplitude term. By taking these two terms as inputs, restoration models are able to handle out-of-sample degradations without fine-tuning. Through extensive experiments across multiple evaluation cases, we demonstrate the effectiveness and flexibility of our proposed framework. Our codes are available at \href{https://github.com/ddghjikle/Out-of-sample-restoration}{Github}.

Figures (10)

• Figure 1: The introduced out-of-sample restoration task is to develop models with the capability of handling unknown degradations. It extends previous restoration researches as paying more attention to cross-degradation generalization.
• Figure 2: Visual examples of the out-of-sample degradation issue. These degraded images are restored by Super-Resolution (SR), Dehaze, SR+Dehaze networks, and our methods. PSNR and SSIM of each example are provided for better comparison.
• Figure 3: Distributions of images suffered from haze, noise, blur, rain, or low-resolution. These visualization results are obtained by leveraging t-SNE van2008visualizing to visualize features generated by the perceptual function johnson2016perceptual.
• Figure 4: Overview of the proposed model reprogramming framework. The input transform module represents input images as amplitude and phase terms. The amplitude term is real-value feature representing image content, and the phase term modulates image styles. A restoration model is used to enhance the amplitude term and align the phase term. Finally, wave functions of processed amplitude and phase are formed and mapped to image space. Dashed rectangles in the top part represent trainable modules.
• Figure 5: Architecture of the Res12 model. Numbers in blue rectangles denote channels of output features.