School of Agricultural Engineering and Food Science, Shandong University of Technology, No. 266 Xincun Xilu, Zibo, Shandong 255049, China; Shandong Provincial Engineering Research Center of Vegetable Safety and Quality Traceability, No. 266 Xincun Xilu, Zibo, Shandong 255049, China; Zibo City Key Laboratory of Agricultural Product Safety Traceability, No. 266 Xincun Xilu, Zibo, Shandong 255049, China.
School of Agricultural Engineering and Food Science, Shandong University of Technology, No. 266 Xincun Xilu, Zibo, Shandong 255049, China.
Int J Food Microbiol. 2024 Oct 2;423:110831. doi: 10.1016/j.ijfoodmicro.2024.110831. Epub 2024 Jul 20.
In this study, a multi-scale attention transformer (MSAT) was coupled with hyperspectral imaging for classifying peanut kernels contaminated with diverse Aspergillus flavus fungi. The results underscored that the MSAT significantly outperformed classic deep learning models, due to its sophisticated multi-scale attention mechanism which enhanced its classification capabilities. The multi-scale attention mechanism was utilized by employing several multi-head attention layers to focus on both fine-scale and broad-scale features. It also integrated a series of scale processing layers to capture features at different resolutions and incorporated a self-attention mechanism to integrate information across different levels. The MSAT model achieved outstanding performance in different classification tasks, particularly in distinguishing healthy peanut kernels from those contaminated with aflatoxigenic fungi, with test accuracy achieving 98.42±0.22%. However, it faced challenges in differentiating peanut kernels contaminated with aflatoxigenic fungi from those with non-aflatoxigenic contamination. Visualization of attention weights explicitly revealed that the MSAT model's multi-scale attention mechanism progressively refined its focus from broad spatial-spectral features to more specialized signatures. Overall, the MSAT model's advanced processing capabilities marked a notable advancement in the field of food quality safety, offering a robust and reliable tool for the rapid and accurate detection of Aspergillus flavus contaminations in food.
在这项研究中，多尺度注意力转换器（MSAT）与高光谱成像相结合，用于分类受不同黄曲霉污染的花生仁。研究结果表明，由于其复杂的多尺度注意力机制增强了分类能力，MSAT 显著优于经典的深度学习模型。多尺度注意力机制通过使用多个多头注意力层来关注细粒度和粗粒度特征来实现。它还集成了一系列尺度处理层，以捕获不同分辨率的特征，并结合自注意力机制来整合不同层次的信息。MSAT 模型在不同的分类任务中表现出色，特别是在区分健康花生仁与黄曲霉污染的花生仁方面，测试准确率达到 98.42±0.22%。然而，它在区分受黄曲霉污染和非黄曲霉污染的花生仁方面面临挑战。注意力权重的可视化明确显示，MSAT 模型的多尺度注意力机制逐渐将其重点从广泛的空间-光谱特征细化到更专业的特征。总体而言，MSAT 模型的先进处理能力标志着食品质量安全领域的显著进步，为快速准确地检测食品中的黄曲霉污染提供了强大可靠的工具。
