Asian Giant Hornets Recognition Using Deep Convolutional Neural Network

Abstract

We trained 2 deep neural networks to identify whether the images with high resolution in the dataset provided by COMAP contain any Asian giant hornet. We divide the classification problem into two subproblems: feature extraction, and image classification. In order to reduce the impact of sample imbalance, we apply image flipping and Borderline-SMOTE methods for data augmentation first, and then divide the data into the training set and the validation set (testing set). Next, we utilize auto-encoder to collect key features of images and the testing loss dips to 0.0274 after training. Next, we establish the Classification-Net to solve the identification problem based on the features just extracted, and the accuracy in testing set reaches 0.8030. Finally, we summarize the main features of having negative labels from three aspects: species characteristics, subject definition and background softness.