Custom convolutional neural network with data augmentation and bayesian optimization for gram-negative bacteria classification

Abstract

One of the newest methods used in image classification is the Convolutional Neural Network. This method uses a large number of hidden layers to process data so that the resulting accuracy is excellent. However, this affects the time of the training process used. The selection of suitable architecture also determines the results of the classification. In this research, the author tries to reduce computational time by reducing the number of layers and using optimization. Transfer learning helps in the preparation of models using pre-trained data before, while data augmentation increases data variation. Bayesian optimization helps to find out momentum values and initial learning rate. The data source of this research is the primary image of Gram-negative bacteria from pneumonia patients. Data was collected at Dr. Soetomo's Microbiology Laboratory in Surabaya, Indonesia. Data distribution includes training, validation, and testing divided by percentage and proportional distribution of the number of files. This research used four classes of Gram-negative bacteria with a total of 1,000 images. An experimental comparison was made with a comparison of the Convolutional Neural Network architecture. The test results show an increase in accuracy by using aiming layers 26-34, having an accuracy range of 99.5% to 99.8%. The computational time required for the training process is around 2 minutes 30 seconds, with a momentum value of 0.92813 and an initial learning level of 0,00022397. The best accuracy errors were obtained at MSE 0.0025, RMSE 0.05, and MAE 0.0025.