Deep Transfer Learning with VGG16 for Insect Pest Recognition Using Visual Features

Abstract

The continuous advancement of artificial intelligence, specifically through Deep Learning techniques using Convolutional Neural Network (CNN) architectures, has profoundly improved the efficiency of image recognition tasks. Nonetheless, classifying natural images of biological subjects like insects remains a difficult endeavor due to high morphological complexity, diverse textures, and hardware-related constraints. This study evaluates the performance of a Transfer Learning approach based on the VGG16 architecture to identify three distinct insect pest species: Henosepilachna vigintioctopunctata, Heteropterus, and Rice Bug. Utilizing a dataset of 1,500 images, the research involves detailed preprocessing steps—including rescaling, contrast adjustment, and image augmentation—to optimize the model's ability to generalize. Experimental findings reveal that the VGG16 model, pre-trained on the ImageNet dataset, achieved a final validation accuracy of 95.80% and a loss value of 0.12 by the 15th epoch. Subsequent evaluations on unseen data demonstrated an overall accuracy of 96% and a consistent F1-score of 0.96 across all classes. Furthermore, confusion matrix analysis verifies a minimal misclassification rate among the species. These results conclude that VGG16-based Transfer Learning is a reliable, robust, and computationally efficient method for insect classification, even when working with restricted dataset sizes.