Fast SAR target recognition based on random convolution features and
ensemble extreme learning machines
Synthetic Aperture Radar (SAR) is one of active microwave imaging sensors, and has been broadly applied in many military and civilian applications, such as disaster surveillance, resource exploration, and missile terminal guidance, etc. It has the features of all-time and all-weather capabilities, multi-bands and multi-polarization modes. The appearance of the target in SAR image is influenced by depression angle of the SAR system and aspect angle of the targets, and the type of noise in SAR image is speckle noise which is a type of multiplicative noise, so how to improve (distorted) target recognition accuracies in standard and extended operating conditions, is still a difficult issue needing to be addressed by SAR Automatic Target Recognition (ATR) researchers. Feature extraction and selection, optimization of classifier design, are two key approaches to improve target recognition rate for SAR ATR. The features extracted from SAR image include gray value, edge, and texture, etc. The state-of-the-art classifiers used in SAR ATR include template matching, SVM, boosting, and sparse representation, etc. Deep convolution neural network has recently demonstrated excellent performance in target detection and recognition tasks. However, lack of training samples and optimization design of deep models are two main issues when applied to SAR target recognition.
“Image Fusion” group, which is from Fundamental Science on Communication Information Transmission and Fusion Technology Laboratory in Hangzhou Dianzi University, proposes an algorithm for SAR target recognition by combination of two dimensional random convolution features and ensemble extreme learning machines. Considering local feature extraction capability based on convolution operations, two dimensional random convolution kernels instead of learned convolution kernels are generated to extract multi-scale local features from SAR images, where kernel widths are firstly randomly selected from the kernel width set, and random kernels with different widths are generated based on uniform distribution. Convolution and square pooling operations are performed in the input image to extract random convolution features, which are then transformed into vectors and combined to form a high dimension feature vector. Secondly, to improve generalization performance of the classifier, random sampling operations based on ensemble learning theory are adopted to perform dimensionality dimension to get a low dimension feature vector, and extreme learning machine(ELM), which has the advantages of fast training speed and few hyperparameters, is used as base classifier. Finally, majority vote method is adopted to combine the classification results of base classifiers to predict target labels. MSTAR database is used to perform SAR target recognition experiments to verify the performance of the proposed algorithm. The hyperparameters which affect recognition performance greatly are firstly analyzed, and it can be concluded that, recognition performance with larger kernel width is higher than that with smaller kernel width, where convolution kernels with small width, such as 3×3, is mostly often used in deep convolution models to perform visible image recognition. Extreme learning machine with small regularization coefficient can achieve good generalization capability and improve recognition performance. SAR target recognition experiments are done under standard operating condition and extended operating conditions, and experimental results demonstrate that, the overall recognition accuracies for ten-class targets with and without distorted configurations are 95.79% and 97.57%, respectively. The proposed algorithm achieves comparable classification performance with deep-learning-based methods which use data augmentation strategy and multiple convolution layers. Meanwhile, the training time has dropped by ten times due to fast training capability of ELM. The proposed algorithm has the advantages of easy implementation and fewer hyperparameters, and improves the generalization performance through adoption of ensemble learning ideas.
“Image Fusion” group from Fundamental Science on Communication Information Transmission and Fusion Technology Laboratory in Hangzhou Dianzi University has fifteen researchers, including six professors, four associate professors, and five lecturers. Their research areas include three aspects: remote image processing, radar-based target detection and tracking, and high-level information fusion. In remote image processing research, they carried out a lot of research work about SAR image processing and interpretation, hyperspectral image classification and anomaly detection, target detection, tracking, recognition, and fusion based on visible and infrared images. They have published more than one hundred journal papers, and obtained more than twenty authorized patents for invention. During the 12th Five-year plan period they were in charge of two sub-projects under The National Basic Research Program (973 Program), one project supported by the State Key Program of National Natural Science of China, and one project supported by the National Key Scientific Instrument and Equipment Development Project. They are now in charge of several projects supported by Chinese Defense Advance Research Program of Science and Technology during the 13th Five-year plan period. Until now they have been honored with one second class National Award for Science and Technology Progress, and two first class Provincial Awards for Science and Technology Progress.
Gu Y, Xu Y. Fast SAR target recognition based on random convolution features and ensemble extreme learning machines[J]. Opto-Electronic Engineering, 2018, 45(1): 170432.