基于扩散变分自编码器的工业缺陷检测算法

Industrial Defect Detection Algorithm Based on Diffusion-Variational Autoencoder

基于重建的检测算法在工业产品自动化缺陷检测中得到了广泛的应用。但是,很多算法重建的图像依然保留了较多的缺陷特征,从而导致检测精度不高。受扩散模型能够建立起噪声和生成图片之间的联系启发,提出了一种基于扩散变分自编码器的工业产品缺陷检测算法。该算法将工业产品中的缺陷视为一种噪声,通过扩散模型的反向去噪过程重建正常图片。在训练阶段,首先利用预训练的基于矢量量化的变分自编码器(Vector Quantized Variational Autoencoder,VQ-VAE)提取工业产品图片的正常特征并添加噪声;然后,利用扩散模型在去噪的过程中消除缺陷特征并保留正常特征,以得到重建的正常图片;最后通过比较重建图片与对应的原始图片之间的差异来判断产品是否存在缺陷并定位缺陷区域。在测试阶段,将输入图片理解为已经添加噪声的图片进行缺陷检测。实验结果表明,该算法的检测精度较其他算法有明显提高。

Reconstruction-based detection algorithms are extensively employed for automatic defect detection in industrial products.Nevertheless,they frequently struggle to effectively eliminate defect features from reconstructed images,resulting in diminished detection accuracy.This paper proposes an industrial product defect detection algorithm based on a diffusion-variational autoencoder to address this issue.The algorithm treats defects as noise and reconstructs typical images through the reverse denoising process of a diffusion model.During training,a pre-trained vector quantized variational autoencoder(VQ-VAE)is utilized to extract normal features and add noise to industrial product images.The diffusion model is then employed to eliminate defect features while preserving normal features during denoising,resulting in reconstructed normal images.Defects can be determined and localized by comparing the reconstructed images with original image.In the testing phase,input images are treated as noise-added for defect detection.Experimental results demonstrate a significant improvement in detection accuracy compared to other algorithms.