Texture suppression in membrane materials based on texture modeling and multi-scale Gabor spatial-frequency fusion

Abstract In membrane structure damage detection, surface textures of membrane materials frequently interfere with damage identification and geometric feature extraction. To address the issues of global detail blurring caused by conventional spatial-domain filtering and the problems of insufficient texture suppression or distortion in non-textured regions induced by frequency-domain filtering due to its poor adaptability to fixed parameters and spectral leakage, this paper proposes a texture suppression method that integrates spatial and frequency domains. First, spatial mathematical models are established for three typical types of membrane textures (quasi-sinusoidal texture, orthogonal grid texture, and diagonal grid texture) to quantitatively analyze their frequency-domain energy distribution characteristics. Based on this analysis, the orientation, scale, and bandwidth parameters of multi-scale Gabor filters are derived, eliminating dependence on manual parameter tuning. Second, a spatial-domain fusion of multi-scale Gabor responses generates a texture energy (TE) map, which integrates with Otsu’s adaptive thresholding and morphological opening–closing operations to precisely segment textured regions and generate optimized masks. Finally, median filtering for suppression is applied only within the textured regions identified by the masks, thereby maximizing the preservation of structural integrity in non-textured regions. Experimental results demonstrate that the proposed texture suppression based on Texture Modeling and multi-scale Gabor spatial-frequency fusion method effectively combines the precision of the frequency domain with the localization advantages of the spatial domain. It achieves significant texture suppression while yielding an average peak signal-to-noise ratio improvement exceeding 10 dB compared to the gradient operator method. When compared to Fourier band-stop filtering, wavelet threshold denoising, and Gaussian filtering methods, it achieves optimal values for both mean gradient magnitude (4.67) and TE (0.03), indicating significantly superior suppression performance over the comparative methods.