Improved Knowledge Distillation for Land-Use Image Classification
Researchers have developed an improved knowledge distillation framework to compress deep convolutional neural networks for land-use image classification. This approach uses a teacher-student learning paradigm where a VGG16 network transfers knowledge to a MobileNetV2 model. By combining hard supervision from ground truth labels with soft supervision using Kullback-Leibler divergence and cosine similarity losses, the method achieved 99.04% accuracy on land-use datasets, outperforming baseline methods while significantly compressing the model. AI