Research On Deep Learning SAR Image Identification Algorithm Based On Limited Labels

Synthetic Aperture Radar(SAR)uses the relative movement between the radar and the target to receive echoes from the detected area for imaging.Compared with optical imaging,SAR imaging systems are not affected by weather,time,etc.,and can detect targets under harsh conditions.Therefore,SAR images are of great value whether they are used in military or civilian applications.It can be applied to natural disaster assessment,landform detection,ship detection,moving target tracking,battlefield monitoring,etc.Therefore,there is a broad research prospect in the automatic recognition of SAR images.In traditional methods,the target recognition of SAR images has encountered a bottleneck.In recent years,the rapid development of deep learning has given new developments in SAR image recognition technology.Recognition models based on deep learning are usually based on supervised learning,which means that the training of the model depends on the input image and the corresponding label of the image.The advantage of supervised deep learning is to extract image features through a large amount of labeled data to distinguish images.When the number of labels decreases,the classification accuracy is limited.However,due to the difficulty of SAR image interpretation,tagged SAR image data is very limited.We solve this problem from two aspects: data augmentation and semi-supervised learning.The main content of this article is as follows:1.Aiming at the problem of insufficient training samples for SAR images in the deep learning process,we found that the data augmentation technology based on motion interpolation can generate any intermediate pictures from any two pictures in the sample to alleviate the lack of data.To improve the accuracy of SAR image recognition.2.By analyzing and comparing the similarities and differences between the motion interpolation-based data amplification technology and the traditional image data amplification methods,we found that the two amplification methods can be used in sequence to amplify the existing labeled data.Experiments show that this method can further improve the performance of CNN model for SAR image recognition.3.In the case of limited label data,the recognition rate of SAR images is not high.We found that we can use semi-supervised learning related methods for SAR image recognition.This method can effectively use unlabeled data,so that the model can learn image feature information in unlabeled data.The recognition performance of the model under the condition of limited labels is greatly improved,which is close to the result of full-label supervised learning.4.We use the attention-related mechanism and the motion interpolation data amplification technology proposed in the previous article to improve the semi-supervised model,which further improves the accuracy of SAR image classification under limited labels.