Abstract
Identifying and classifying bacterial species from microscopic images is crucial for medical applications like prevention, diagnosis, and treatment. However, because of their diversity and variability in appearance, manually classifying bacteria is difficult and time-consuming. This work suggests employing deep learning architecture to automatically categorize bacterial species in order to overcome these difficulties and raise the accuracy of bacterial species recognition. We have evaluated our suggested approach using the Digital Images of Bacteria Species (DIBaS), a publicly accessible resource of photographs of tiny bacteria. This work uses a dataset that differs in terms of bacterial morphology, staining methods, and imaging circumstances. This paper aims to enhance the accuracy and reduce the computational requirements for Convolutional Neural Networks (CNN) based classification of bacterial species using GoogLeNet and AlexNet to train the models. This paper focuses on employing transfer learning to retrain pre-trained CNN models using a dataset consisting of 2000 images encompassing 12 distinct bacteria species known to be harmful to human health. The concept of transfer learning was utilized to expedite the network's training process and enhance its categorization performance. The results are promising, with the method achieving an accuracy of 98.7% precision, recall of 99.50%, and an F1-score of 99.45% with classifier speed. Furthermore, the proposed bacteria classification approach demonstrated strong performance, irrespective of the size of the training data used. This paper contributes by automating bacterial classification to facilitate faster and more accurate identification of bacterial species, which facilitates the treatment of infections and related diseases, in addition to monitoring public health, and promoting the wise use of antimicrobial drugs. To improve outcomes in the future, researchers can also integrate deep learning techniques with other machine learning methods.