From Spectrum to Image: A Novel Deep Clustering Network for Lactose-Free Milk Adulteration Detection

Abstract

Traditional clustering methods are often ineffective in extracting relevant features from high-dimensional, nonlinear near-infrared (NIR) spectra, resulting in poor accuracy of detecting lactose-free milk adulteration. In this paper, we introduce a clustering model based on Gram angular field and convolutional depth manifold (GAF-ConvDuc). The Gram angular field accentuates variations in spectral absorption peaks, while convolution depth manifold clustering captures local features between adjacent wavelengths, reducing the influence of noise and enhancing clustering accuracy. Experiments were performed on samples from 2250 milk spectra using the GAF-ConvDuc model. Compared to K-means, the silhouette coefficient (SC) increased from 0.109 to 0.571, standardized mutual information index (NMI) increased from 0.696 to 0.921, the Adjusted Randindex (ARI) increased from 0.543 to 0.836, and accuracy (ACC) increased from 67.2% to 88.9%. Experimental results indicate that our method is superior to K-means, Variational Autoencoder (VAE) clustering, and other approaches. Without requiring pre-labeled data, the model achieves higher inter-cluster separation and more distinct clustering boundaries. These findings offer a robust solution for detecting lactose-free milk adulteration, crucial for food safety oversight.