Aflatoxins are the most dangerous mycotoxin produced by Aspergillus fungal species, such as Aspergillus flavus and Aspergillus parasiticus, and can cause various health problems, including liver cancer, in humans. Figs and many agricultural products are affected by aflatoxins. It is crucial to detect it before consumption since it is impossible to clean aflatoxin from contaminated foods. Chromatographic methods are considered the gold standard for aflatoxin detection, but these methods are pretty expensive, time-consuming, and destructive. Therefore, various studies have been conducted on non-invasive aflatoxin detection using optical spectroscopic methods. Specifically, aflatoxin-contaminated figs are sorted manually by employees in production facilities using the Bright Greenish Yellow Fluorescence (BGYF) method under ultraviolet (UV) light. However, accurate and safe manual sorting depends on expertise of employees, along with this long exposure time to UV radiation may cause employees skin cancer and eye disorders. This study presents a deep transfer learning-based approach for non-invasive detection and classification of aflatoxin-contaminated dried figs using images captured under UV light. Pre-trained transfer learning models, such as DenseNet, ResNet, VGG, and InceptionNet, are applied to the dataset, but the accuracy of these models does not outperform the other methods that detect aflatoxin with the BGYF method. Therefore, fine-tuning is performed on the models. As a result, training accuracy of 98.57% and validation accuracy of 97.50% is obtained using the DenseNet169 model. The experimental results show that our proposed method achieves the highest accuracy among other methods. Also, the proposed method shows that deep CNN can be used to automatically, rapidly, and effectively detect aflatoxin-contaminated figs.
Read full abstract