Image Enhancement Utilizing Generative Adversarial Networks for Improving the Accuracy of Pest Insect Classification

The accuracy of automated insect pest identification systems has been significantly improved with the advent of models based on deep neural networks in machine vision. To build a robust deep neural network categorization model, however, a large quantity of image data is required. It may be time-consuming and labor-intensive to collect and annotate massive datasets while creating image data. The purpose of this research is to cut down on the effort and time needed to do tasks like these. Insects' pest convolutional neural network (CNN) classification models are improved in this study by using a generative adversarial network (GAN) to synthesize training pictures from sticky paper trap photographs. uniform with little data, GAN was able to synthesis pictures to create a more uniform distribution and a larger pool of training images. When evaluating the quality of the synthetic pictures, t-SNE visualization was used as a test. The minimal needed number of pictures for applying GAN-based augmentation was calculated in order to test the constructed network. Classifier models for insect pests were also tested and compared across three augmentation scenarios: no augmentation, traditional augmentation, and GAN-based augmentation. Classifier models trained using GAN -enhanced photographs outperformed those trained with typically augmented photos on tests (F1-score, 0.95 vs. 0.92 respectively). This study demonstrates that insect-pest CNN classification algorithms may improve in performance by employing a GAN-based augmentation strategy instead of a traditional one, while also requiring less data gathering effort. When just a small number of images is available, the suggested method may be simply applied to a wide range of agricultural applications.