Transfer learning-based knowledge distillation for UAV disaster detection

Traditional image classification methods have low accuracy, and deep learning-based classification methods have high accuracy, but they are computationally intensive and costly, and mobile devices relying on limited arithmetic power find it difficult to promote their use in small-sample applications such as UAVs. And in UAV disaster detection, it is necessary to simultaneously ensure the high accuracy of images and lightweight deployment requirements. The knowledge refinement technique based on transfer learning synthesizes the advantages of the two methods and becomes the key to solving this contradiction. In order to improve the accuracy of the model, a convolutional neural network structure based on migration learning is established, which utilizes the Xception model trained by migration learning on a preprocessed dataset, and then the "black knowledge" of Xception is fed back to the lightweight student model through the knowledge refinement technique. By comparing several student models such as EfficientNet, NASNetMobile, and MobileNet, the MobileNet model with the best results was selected, thus ensuring the accuracy of UAV disaster detection by further improving the accuracy and lightness of image classification in small sample applications.