Department of Environmental Engineering and Science, Feng Chia University, 100, Wenhwa Rd., Seatwen, Taichung, 40724, Taiwan.
Bachelor Program in Interdisciplinary Studies, National Yunlin University of Science and Technology, 123 University Road, Section 3, Douliou, Yunlin, 64002, Taiwan.
Environ Sci Pollut Res Int. 2021 Jun;28(24):31920-31932. doi: 10.1007/s11356-021-13010-9. Epub 2021 Feb 22.
Rapid environmental microorganism (EM) classification under microscopic images would help considerably identify water quality. Because of the development of artificial intelligence, a deep convolutional neural network (CNN) has become a major solution for image classification. Three popular CNNs, referred to as ResNet50, Vgg16, and Inception-v3, were transferred to identify the EM images present on the Environmental Microorganism Dataset (EMDS), and EMAD was the small dataset, which only has 294 EM images with 21 EM classes. Besides data augmentation, optimizing the fully connected layer of CNN, i.e., both optimally fine-tuned neuron number and dropout rate, was adopted to enhance the performance produced by CNN. The discussions on the causes of the accuracy improved by optimization are also provided. The results showed that the Inception-v3 model obtained 84.9% of the accuracy and performed better than the other two famous CNNs. Also, the implement of data augmentation enhanced the performance of Inception-v3 on EMDS. To add to that, the optimized Inception-v3 model archived 90.5% of the accuracy, and this result demonstrated the improvement effect obtained by using genetic algorithm (GA) to optimize the fully connected layer of the Inception-v3. Therefore, the optimize Inception-v3 with data augmentation process obtained the accuracy of 92.9% and improved almost 21% higher than that obtained from the famous Vgg16. In addition, the optimized Inception-v3 would need less neurons, when compared with that of the optimized Vgg16 possibly. This optimized Inception-v3 could provide a solution to the EM classification in microscope with a digital camera system.
快速的环境微生物（EM）分类在微观图像下有助于极大地识别水质。由于人工智能的发展，深度卷积神经网络（CNN）已成为图像分类的主要解决方案。三种流行的 CNN，分别称为 ResNet50、Vgg16 和 Inception-v3，被转移到识别环境微生物数据集（EMDS）中存在的 EM 图像，而 EMAD 是一个小数据集，只有 294 张 EM 图像，分为 21 个 EM 类别。除了数据增强，优化 CNN 的全连接层，即优化微调神经元数量和辍学率，也被采用来提高 CNN 产生的性能。还提供了对优化导致的准确性提高的原因的讨论。结果表明，Inception-v3 模型获得了 84.9%的准确性，比另外两种著名的 CNN 表现更好。此外，数据增强的实现增强了 Inception-v3 在 EMDS 上的性能。除此之外，优化的 Inception-v3 模型获得了 90.5%的准确性，这一结果证明了使用遗传算法（GA）优化 Inception-v3 的全连接层所获得的改进效果。因此，经过数据增强和优化的 Inception-v3 获得了 92.9%的准确性，比著名的 Vgg16 提高了近 21%。此外，优化的 Inception-v3 可能需要比优化的 Vgg16 更少的神经元。这个优化的 Inception-v3 可以为带有数码相机系统的显微镜中的 EM 分类提供解决方案。
