Enhancing Faster-RCNN Performance: A Hybrid Approach Integrating VAE, HOG, and Recursive Feature Elimination

Abstract

Deep learning has transformed image analysis by providing powerful feature extraction techniques, particularly beneficial for detecting microscopic pathogens such as Vibrio parahaemolyticus. Traditional methods often fail to capture the complex, high-dimensional features required for accurate bacterial classification. This study presents the enhancement of Faster-RCNN performance through advanced feature extraction techniques for the classification of Vibrio parahaemolyticus bacteria. The goal is to improve classification accuracy by integrating sophisticated feature extraction methods. Microscopic images observed directly are initially processed using the ResNet architecture to derive preliminary features. Variational Autoencoders (VAE) are then employed to extract high-dimensional, abstract features, while Histogram of Oriented Gradients (HOG) captures shape and orientation based features. Recursive Feature Elimination (RFE) is used to optimize these feature sets, leading to significant improvements in classification performance. The VAE+RFE + Faster-RCNN approach achieves a detection accuracy of 92%, precision of 88%, recall of 93%, and an F1-score of 90%. In comparison, the HOG + RFE + Faster-RCNN configuration results in 85% accuracy, 82% precision, 87% recall, and an F1-score of 84%. The conventional Faster-RCNN model records an accuracy of 89%, precision of 86%, recall of 90%, and an F1-score of 88%. These results underscore the substantial impact of advanced feature extraction techniques on optimizing Faster-RCNN performance, demonstrating significant improvements in classification accuracy and efficiency.