A Study of Roughness Detection Based on Attention and Multi-Spectral Laser Speckle Images

The roughness of a machined metal surface is critical to the performance and service life of the work-piece. With the development of machine vision technology in the field of roughness inspection, laser speckle measurement has been widely used. However, the single-wavelength laser speckle-based detection technology is difficult to comprehensively characterize the complex surface topography due to the single spectral information, resulting in limited detection accuracy. To address the above problems, this paper proposes a roughness detection method based on attention and multi-spectral laser speckle images. The method extracts the salient features of each wavelength band by using parallel channel attention through the multi-spectral convolutional attention module and realizes crosswavelength fusion and interaction achieved through shared spatial attention and skip connections; and computes the selfattention weights in the spectral dimension through the multispectral self-attention module, which dynamically enhances the key wavelength contribution and suppresses the redundant information. Experiments show that the multi-spectral fusion strategy effectively overcomes the limitations of singlewavelength information, and the algorithm proposed in this paper has a detection accuracy of 98% on the test set, which is 4.4% higher than the single-wavelength baseline model ConvNeXt-Tiny.