An unsupervised denoising via differential feature learning under high-level noise indistinguishable from defect edges

Abstract

In industrial inspection scenarios, the superposition of multiple acoustic sources and the nonlinear, time-varying characteristics of noise significantly complicate the direct characterization of defects from raw detection signals. In the absence of accessible paired training data, a critical challenge lies in simultaneously suppressing noise while maximally preserving salient signals and recovering defect features obscured by high-level noise. To address this, An unsupervised denoising network is proposed, incorporating two principal innovations. First, to achieve both high-level noise suppression and the recovery of noise-obscured defect regions, a generator architecture based on a degradation-decoupling mechanism is designed. A decoupled attention module is proposed to enhance the separability among defect bodies, noise-masked defect areas, and high-noise backgrounds. By leveraging their complementary structural features, more accurate reconstruction of complete defect morphology is enabled. Moreover, the high perceptual similarity between defect edges and noise, coupled with the insufficient domain constraints of unsupervised learning, tends to cause structural distortion of defects. To address this issue, A differential feature modeling strategy for regions of interest is proposed to improve the preservation of defect-relevant information in regions of interest. Experimental results demonstrate that the proposed unsupervised denoising network effectively suppresses noise while preserving signals of interest and recovering defect features obscured by interference. Comparative evaluations further validate its superior overall performance over both supervised and unsupervised deep learning-based denoising approaches.