Multi-class Pathogenic Microbes Classification by Stochastic Gradient Descent and Discriminative Fine-Tuning on Different CNN Architectures

The detection of microorganisms is an important task in the clinical microbiology field. It is equally important during the pandemic breakout. Pathogenic microbes’ orientational behavior helps in distinguishing them. However, it is not an easy task to classify them based on that behavior only. In this research work, image processing and CNN methods like Resnet50, DenseNet121, Inception-ResNetv2, and MobileNetv2 have been implemented to classify species of 33 different pathogenic microbes. The pathogenic microbes have stained with Gramm’s method to distinguish them as gram-positive and gram-negative bacteria. Lactobacillus, Staphylococcus, and Enterococcus are used for their intra-general classification. Further, Stochastic gradient descent and fine-tuning are used to tune the learning rate. The result shows that 90.62 accuracies have been obtained using ResNet architecture for discriminative fine-tuning (DFT) and 92.96 accuracies have been obtained using stochastic gradient descent with the warm restarts (SGDR) approach. Similarly, 91.41 accuracies have been obtained using DenseNet 121 architecture for DFT and an accuracy of 98.7 has been obtained using Stochastic gradient descent with warm restarts approach. Also, an accuracy of 96.88 has been obtained using MobileNet architecture for DFT and an accuracy of 99.2 using Stochastic gradient descent with the warm restarts method. Further, Inception-ResNetV2 architecture has obtained an accuracy of 99.15 using DFT and 99.53 for the SGDR approach.