A Novel Lightweight Sarnet With Clock-Wise Data Amplification for Sar Atr

Abstract

Convolutional Neural Network (CNN) models applied to synthetic aperture radar automatic target recognition (SAR ATR) universally focus on two important issues: overfitting caused by lack of sufficient training data and independent variations like worse estimates of the aspect angle, etc. To this end, we developed a lightweight CNN-based method named SARNet to accomplish the classification task. Firstly, a clockwise data amplification approach is presented to generate adequate SAR images without requiring many raw SAR images, effectively avoiding overfitting in the course of training. Then a SARNet is devised to process the extracted features from SAR target images and work on classification tasks with parameters fine-tuning under comparative models. To enhance and structurally organize the representation of learned proposed model, various activation functions are explored in this paper. Furthermore, due to the pioneering conducted experiments, training samples in the MSTAR and extended MSTAR database are utilized to demonstrate the robustness and effectiveness of the lightweight model. Experimental results have shown that our proposed model has achieved a 98.30% state-of-the-art accuracy.

Robust SAR Automatic Target Recognition Based on Transferred MS-CNN with L2-Regularization

“…AF-CNN [44] is an additional feature-based CNN architecture which does not need additional preprocessing process or pose information, which boosts 4.38% improvement compared with SVM. e traditional CNN [ Method Accuracy (%) SVM [41] 90.00 AlexNet 93.55 AE&LSVM [42] 94.14 DNPP-L1 [43] 94.14 AF-CNN [44] 94.38 Gabor & LPQ & ELM [45] 94.80 CNN [46] 95.90 Unsupervised K-means & data amplification [47] 96.67 LeNet [48] 97.29 ResNet-50 [24] 97.66 SARNet [49] 98.30 JLSND&SRC 1 [50] 98.30 Proposed 98.93…”

Synthetic Aperture Radar SAR Image Target Recognition Algorithm Based on Attention Mechanism

“…There are various methods to reduce the weight of the model, such as adding an attention mechanism, network pruning [20], and separable convolution [21]. For example, in [22], in 2019, Zhao et al proposed a lightweight CNN model to avoid the over fitting problem caused by the lack of data, and achieved 98.30% accuracy on MSTAR data set. In [23], Ying used selfattention and knowledge distillation to achieve weight reduction of the model, and good results were achieved on the MSTAR dataset.…”