Research And Implementation Of Sample Amplification Technology Based On Deep Learning

Nowadays,in the era of big data,related research on deep learning is in full swing.A sufficient number of training samples are the premise and foundation of neural network model training.At the same time,the spatial distribution and diversity of sample data guarantee the robustness and generalization ability of network model,but in the actual production and under the background of big data,the cost of acquiring sample data is high,or the acquired data has a large number of repetitions,or the existence of required samples has a low probability of occurrence,which leads to a very limited number of useful sample data,namely the problem of "big data,small sample",and the field of remote sensing images also faces the same challenge,so the research on sample amplification technology is of great significance.In view of the above background,this thesis studies the sample amplification technology based on deep learning to solve the problem of the limited number of samples in the field of remote sensing images.The main work content includes:1.Evaluate the traditional existing sample amplification technology,mainly including the sample amplification technology based on texture modeling,the sample amplification technology based on the Cycle GAN network and the sample amplification technology of the improved Cycle GAN network,and analyze the network structure and the design of loss function.At the same time,tests were carried out on the official data set and remote sensing data set respectively,and the existing problems were analyzed according to the test results.2.In view of the imaging characteristics of remote sensing images and the limitations of traditional sample amplification techniques,a combination of semantic segmentation and generation antagonism is proposed for data sample amplification.The core idea is that the semantic segmentation network is used to segment and extract the target,and then the antagonistic network is combined with the semantic segmentation network,and finally the sample is generated and the background is fused to achieve sample amplification.3.According to the existing sample quality assessment methods,design the sample quality evaluation of the target detection.The target detection framework is based on an improved version of the YOLO model,which can balance detection accuracy and speed on Synthetic Aperture Radar(SAR)data samples.For sample quality evaluation,it simulates the scarce scene of real remote sensing images,by adding different numbers of generated samples,compare the target detection performance indicators of real remote sensing images and amplified training sets,and evaluate the quality of the generated sample data.