| With the rapid development of Synthetic Aperture Radar(SAR)technology,high-resolution SAR images are broadly used in military and civilian fields,and play an irreplaceable role in military reconnaissance,resource exploration,environmental monitoring and other tasks.In recent years,the demand for SAR image data has increased.Faced with the massive SAR image data provided by aerospace sensors,there is an urgent need to develop accurate,efficient,and generalized SAR image interpretation technology.Due to the complexity and diversity of SAR images,traditional target detection and recognition methods are difficult to meet the needs of practical applications since they are cumbersome,low accuracy,and poor adaptability.In this context,SAR image object detection and recognition methods are studied via deep learning in this dissertation,aiming to further improve the performance of SAR image object detection and recognition.The specific research work is as follows:In the aspect of SAR image object detection,the performance of several deep learning object detection algorithms are analyzed,according to the evaluation indicators of MS COCO and PASCAL VOC.In views of the characteristics of SAR image objects,YOLOv3 is selected as the basis network model.To solve the problems of missed detection of small objects in SAR image object detection and poor distinguishing ability in complex environments,an improved SAR image object detection method based on the YOLOv3 model is proposed.The model is optimized in the aspects of anchor frame size,network structure,and loss function.Firstly,the size of the anchor frame is redesigned according to the statistical analysis of the SAR image data set.Secondly,the structure of the model is optimized through redesigning the residual unit,simplifying the network model,increasing the information and prediction scale of shallow feature map,and using ASFF to construct the feature pyramid.Finally,the loss function is optimized by setting the weighting coefficient to balance the weight of small object.The performance of the proposed model is verified on the SSDD dataset and HRSID dataset.In comparison with the original YOLOv3 model,the average accuracy of the improved model is increased by 9.28% and 12.92% respectively,F1 scores is increased by 6% and7% respectively.The detection ability of small objects and the distinguishing effect in complex environments are improved.Hence,it may be concluded that the proposed SAR image object detection method is effective.In the aspect of SAR image object recognition,the object recognition method with strong adaptability is been explored in this context.Firstly,two deep learning target recognition models based on convolutional neural network(CNN)and stacked auto encoder(SAE)are constructed,respectively.To solve the problem of small samples,the experimental data set is constructed and expanded through using electromagnetic simulation data,moving,and increasing noise.The two models are verified in the object recongnition experiments of BTR70,BMP2,and T72.Experimental results show that the recognition model based on CNN has high recognition accuracy,but is sensitive to noise.The recognition model based on SAE has a slightly lower recognition accuracy but is adaptable to noise.Based on the characteristics of the two models,a SAR image target recognition model based on deep feature fusion is proposed.The model combines two deep features extracted by CNN and SAE network.The experimental results show that the proposed target recognition method not only has high recognition accuracy,but also has anti-interference ability against noise.In all,the proposed method has good comprehensive performance. |
