Freezing pre-trained parameters of encoders for denoisers: Expanding pixel involvement and filtering out high-frequency noise.

Abstract

Image denoising is a pivotal and intricate task within the field of deep learning and the challenge of generalization poses significant difficulties for many denoisers when faced with out-of-distribution (OOD) noise. Model generalization performance depends critically on three components: model architectural design; dataset composition; and carefully designed training methods. While the first two factors have consistently been the focus of research, there remains a significant gap in the exploration of training methods. In this paper, we prove a straightforward and efficient training strategy to improve the generalization of denoisers. Specifically, we propose that freezing the encoders' pre-trained parameters and updating the decoders' parameters during the training process. Experimental results confirm that our proposed method, despite its simplicity, achieves more robust denoising performance across various noise types. This demonstrates the effectiveness of our training strategy in enhancing network generalization. To the best of our knowledge, we are the first to demonstrate that freezing pre-trained parameters of encoders during the training process not only expands the range of input pixels that strongly influence the denoising results but also effectively filters out high-frequency noise signals, thereby improving the performance of the denoiser.