Remote Sensing Image Classification Network Based On Knowledge Distillation And Multi-instances Learning

With the continuous development of remote sensing technology,high-resolution remote sensing image data brings new opportunities for remote sensing image classification,but it also brings challenges.Because remote sensing image data usually has the following characteristics: there are many objects in the image,complex background information,non-centralized distribution of key objects,and irregular deformation of objects.Therefore,remote sensing image classification is facing great difficulties.Nowadays,under the background of deep learning,there are extensive research on applying deep learning to remote sensing image scene classification,which promotes the development of remote sensing image scene classification methods.At the same time,with the development of scene classification methods,the size of the model and the amount of parameters are also increased,which leads to the consumption of a lot of time and computing resources in the process of model training and inference.Therefore,how to compress the model size without affecting the accuracy is also a current hot research topic.In response to the above problems,this thesis introduces knowledge distillation and multi-instance learning framework.Related work can be summarized as follows:(1)This thesis proposes a remote sensing image classification method based on knowledge distillation multi-instance learning.By introducing knowledge distillation and compressing the network,the problems of large computing resource consumption and time-consuming reasoning are effectively solved.For the irregular deformation of objects,an attention-based multi-instance learning is proposed to obtain better local semantic information.At the same time,the distillation loss and classification loss are jointly optimized to improve the overall network classification performance.(2)On the basis of knowledge distillation,this thesis constructs a remote sensing scene classification network based on self-distillation and multi-instance learning.Aiming at the problem that the image consists of many objects and complex background information,by introducing multi-instance learning,the feature extraction is enhanced by self-distillation to obtain better example representations and help the network learn the position weight information of different semantic examples.At the same time,it solves the problems of large computing resource consumption and time-consuming reasoning.Finally,the distillation loss,instance-level loss and packet-level loss are jointly optimized to improve the overall network classification performance.This thesis proves that the method has competitive performance in remote sensing image classification tasks on three public remote sensing datasets UCM,AID and NWPURESISC.Compared with other methods,the method in this thesis improves the performance of remote sensing scene classification while reducing the number of parameters and calculations of the model.In addition,experiments on natural scene CIFAR-100 demonstrate the effectiveness of the method.