自監(jiān)督解耦動(dòng)態(tài)分類器的小樣本類增量SAR圖像目標(biāo)識(shí)別
doi: 10.11999/JEIT231470
-
國(guó)防科技大學(xué)電子科學(xué)學(xué)院 長(zhǎng)沙 410073
Few-Shot Class-Incremental SAR Image Target Recognition using Self-supervised Decoupled Dynamic Classifier
-
College of Electronic Science and Technology, National University of Defense Technology, Changsha 410073, China
-
摘要: 為提升基于深度學(xué)習(xí)(DL)的合成孔徑雷達(dá)自動(dòng)目標(biāo)識(shí)別(SAR ATR)系統(tǒng)在開放動(dòng)態(tài)的非合作場(chǎng)景中對(duì)新類別目標(biāo)的持續(xù)敏捷識(shí)別能力,該文研究了SAR ATR的小樣本類增量學(xué)習(xí)(FSCIL)問題,并提出了自監(jiān)督解耦動(dòng)態(tài)分類器(SDDC)。針對(duì)FSCIL 中“災(zāi)難性遺忘”和“過擬合”本質(zhì)難點(diǎn)和SAR ATR領(lǐng)域挑戰(zhàn),根據(jù)SAR圖像目標(biāo)信息的部件化與方位角敏感性特點(diǎn),于圖像域構(gòu)建了基于散射部件混淆與旋轉(zhuǎn)模塊(SCMR)的自監(jiān)督學(xué)習(xí)任務(wù),以提升目標(biāo)表征的泛化性與穩(wěn)健性。同時(shí),設(shè)計(jì)了類印記交叉熵(CI-CE)損失并以參數(shù)解耦學(xué)習(xí)(PDL)策略對(duì)模型動(dòng)態(tài)微調(diào),以對(duì)新舊知識(shí)平衡判別。實(shí)驗(yàn)在由MSTAR和SAR-AIRcraft-1.0數(shù)據(jù)集分別構(gòu)建的覆蓋多種目標(biāo)類別、觀測(cè)條件和成像平臺(tái)的FSCIL場(chǎng)景上驗(yàn)證了該算法開放動(dòng)態(tài)環(huán)境的適應(yīng)能力。
-
關(guān)鍵詞:
- SAR目標(biāo)識(shí)別 /
- 小樣本類增量學(xué)習(xí) /
- 自監(jiān)督學(xué)習(xí) /
- 深度學(xué)習(xí)
Abstract: To power Deep-Learning (DL) based Synthetic Aperture Radar Automatic Target Recognition (SAR ATR) systems with the capability of learning new-class targets incrementally and rapidly in openly dynamic non-cooperative situations, the problem of Few-Shot Class-Incremental Learning (FSCIL) of SAR ATR is researched and a Self-supervised Decoupled Dynamic Classifier (SDDC) is proposed. Considering solving both the intrinsic Catastrophic forgetting and Overfitting dilemma of the FSCIL and domain challenges of SAR ATR, a self-supervised learning task powered by Scattering Component Mixup and Rotation (SCMR) is designed to improve the model’s generalizability and stability for target representation, leveraged by the partiality and azimuth dependence of target information in SAR imagery. Meanwhile, a Class-Imprinting Cross-Entropy (CI-CE) and a Parameter Decoupled Learning (PDL) strategy are designed to fine-tune networks dynamically to identify old and new targets evenly. Experiments on various FSCIL scenarios constructed by the MSTAR and the SAR-AIRcraft-1.0 datasets covering diverse target categories, observing environments, and imaging payloads, verify the method’s adaptability to openly dynamic world. -
1 散射混淆操作(SMO)處理過程
輸入:小批次基類訓(xùn)練樣本 ${\boldsymbol} = \{ ({{\boldsymbol{x}}_i},{y_i})\} _{i = 1}^{|{N_{}}|} \in {{\boldsymbol{D}}^1}$ 初始化:混淆次數(shù) R ,混淆樣本存儲(chǔ)列表${S_{{\text{smo}}}}$ For r = 1 to R do 步驟1 采樣隨機(jī)順序數(shù)據(jù)${\boldsymbol{\tilde b}} = \{ ({{\boldsymbol{\tilde x}}_i},{\tilde y_i})\} _{i = 1}^{|{N_{}}|}$ 步驟2 獲取原始$\{ {y_i}\} _{i = 1}^{|{N_{}}|}$與隨機(jī)順序$\{ {\tilde y_i}\} _{i = 1}^{|{N_{}}|}$標(biāo)簽 步驟3 從${\boldsymbol}$中選取${\boldsymbol}' = \{ ({{\boldsymbol{x}}'_i},{y'_i})\} $,從${\boldsymbol{\tilde b}}$中選取
$ {\boldsymbol{\tilde b}}' = \{ ({{\boldsymbol{\tilde x}}'_i},{\tilde y'_i})\} $,滿足${y'_j} \ne {\tilde y'_j},\forall j \in |{N_{}}|$步驟4 根據(jù)式(1)從${\boldsymbol}'$與 $ {\boldsymbol{\tilde b}}' $中生成散射混淆樣本集
$ {{\stackrel \frown{{\boldsymbol}} }} = \{ ({{{\stackrel \frown{{\boldsymbol{x}}} }}_i},{\stackrel \frown{y} _i})\} $步驟5 將$ {{\stackrel \frown{{\boldsymbol}} }} $添加至${S_{{\text{smo}}}}$ End for 輸出:${S_{{\text{smo}}}}$ 下載: 導(dǎo)出CSV
表 1 MSTAR-FSCIL數(shù)據(jù)集配置
學(xué)習(xí)階段 序號(hào) 類別 型號(hào) 訓(xùn)練集 測(cè)試集 基類 1 BTR70 c71 233 196 2 2S1 b01 299 274 3 BRDM2 E-71 298 274 4 BMP2 9563 233 196 增量類 5 ZIL131 E12 5 274 6 T62 A51 5 274 7 D7 13015 5 274 8 BTR60 7532 5 195 9 T72 132 5 196 10 ZSU234 d08 5 274 下載: 導(dǎo)出CSV
表 2 SAR-AIRcraft-1.0-FSCIL數(shù)據(jù)集配置表
學(xué)習(xí)階段 序號(hào) 類別 訓(xùn)練集 測(cè)試集 基類 1 Other 2 000 200 2 A220 2 000 200 3 Boeing787 2 000 200 4 Boeing737 2 000 200 增量類 5 A320 5 200 6 ARJ21 5 200 7 A330 5 200 下載: 導(dǎo)出CSV
表 3 SDDC算法模塊與損失貢獻(xiàn)(%)
遷移性 判別性 Avg.Acc Acc.Old Acc.New PD SMO SRO CI-CE PDL 75.26 78.87 49.02 41.50 √ 78.35 81.41 54.16 37.77 √ 77.79 81.27 52.24 35.29 √ √ 79.50 82.68 54.46 34.47 √ √ √ 73.30 72.62 79.12 55.83 √ √ √ 75.81 79.01 47.22 42.35 √ √ √ √ 81.73 83.37 66.70 31.61 下載: 導(dǎo)出CSV
-
[1] 張路, 廖明生, 董杰, 等. 基于時(shí)間序列InSAR分析的西部山區(qū)滑坡災(zāi)害隱患早期識(shí)別——以四川丹巴為例[J]. 武漢大學(xué)學(xué)報(bào): 信息科學(xué)版, 2018, 43(12): 2039–2049. doi: 10.13203/j.whugis20180181.ZHANG Lu, LIAO Mingsheng, DONG Jie, et al. Early detection of landslide hazards in mountainous areas of West China using time series SAR interferometry-A case study of Danba, Sichuan[J]. Geomatics and Information Science of Wuhan University, 2018, 43(12): 2039–2049. doi: 10.13203/j.whugis20180181. [2] 李永禎, 黃大通, 邢世其, 等. 合成孔徑雷達(dá)干擾技術(shù)研究綜述[J]. 雷達(dá)學(xué)報(bào), 2020, 9(5): 753–764. doi: 10.12000/JR20087.LI Yongzhen, HUANG Datong, XING Shiqi, et al. A review of synthetic aperture radar jamming technique[J]. Journal of Radars, 2020, 9(5): 753–764. doi: 10.12000/JR20087. [3] 傅興玉, 尤紅建, 付琨. 基于鄰域均方連續(xù)差分的SAR圖像邊緣提取算法[J]. 電子與信息學(xué)報(bào), 2012, 34(5): 1030–1037. doi: 10.3724/SP.J.1146.2011.00920.FU Xingyu, YOU Hongjian, and FU Kun. An approach to extract edge in SAR image based on square successive difference of neighborhood averages[J]. Journal of Electronics & Information Technology, 2012, 34(5): 1030–1037. doi: 10.3724/SP.J.1146.2011.00920. [4] 羅汝, 趙凌君, 何奇山, 等. SAR圖像飛機(jī)目標(biāo)智能檢測(cè)識(shí)別技術(shù)研究進(jìn)展與展望[J]. 雷達(dá)學(xué)報(bào), 2024, 13(2): 307–330. doi: 10.12000/JR23056.LUO Ru, ZHAO Lingjun, HE Qishan, et al. Intelligent technology for aircraft detection and recognition through SAR imagery: Advancements and prospects[J]. Journal of Radars, 2024, 13(2): 307–330. doi: 10.12000/JR23056. [5] 周大蔚, 汪福運(yùn), 葉翰嘉, 等. 基于深度學(xué)習(xí)的類別增量學(xué)習(xí)算法綜述[J]. 計(jì)算機(jī)學(xué)報(bào), 2023, 46(8): 1577–1605. doi: 10.11897/SP.J.1016.2023.01577.ZHOU Dawei, WANG Fuyun, YE Hanjia, et al. Deep learning for class-incremental learning: A survey[J]. Chinese Journal of Computers, 2023, 46(8): 1577–1605. doi: 10.11897/SP.J.1016.2023.01577. [6] 丁柏圓, 文貢堅(jiān), 余連生, 等. 屬性散射中心匹配及其在SAR目標(biāo)識(shí)別中的應(yīng)用[J]. 雷達(dá)學(xué)報(bào), 2017, 6(2): 157–166. doi: 10.12000/JR16104.DING Baiyuan, WEN Gongjian, YU Liansheng, et al. Matching of attributed scattering center and its application to synthetic aperture radar automatic target recognition[J]. Journal of Radars, 2017, 6(2): 157–166. doi: 10.12000/JR16104. [7] HUMMEL R. Model-based ATR using synthetic aperture radar[C]. The IEEE 2000 International Radar Conference, Alexandria, USA, 2000: 856–861. doi: 10.1109/RADAR.2000.851947. [8] CHEN Sizhe, WANG Haipeng, XU Feng, et al. Target classification using the deep convolutional networks for SAR images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2016, 54(8): 4806–4817. doi: 10.1109/TGRS.2016.2551720. [9] REBUFFI S A, KOLESNIKOV A, SPERL G, et al. iCaRL: Incremental classifier and representation learning[C]. 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 5533–5542. doi: 10.1109/CVPR.2017.587. [10] HOU Saihui, PAN Xinyu, LOY C C, et al. Learning a unified classifier incrementally via rebalancing[C]. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, USA, 2019: 831–839. doi: 10.1109/CVPR.2019.00092. [11] TAO Xiaoyu, HONG Xiaopeng, CHANG Xinyuan, et al. Few-shot class-incremental learning[C]. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2020: 12180–12189. doi: 10.1109/cvpr42600.2020.01220. [12] ZHANG Chi, SONG Nan, LIN Guosheng, et al. Few-shot incremental learning with continually evolved classifiers[C]. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA, 2021: 12450–12459. doi: 10.1109/CVPR46437.2021.01227. [13] PENG Can, ZHAO Kun, WANG Tianren, et al. Few-shot class-incremental learning from an open-set perspective[C]. 17th European Conference on Computer Vision, Tel Aviv, Israel, 2022: 382–397. doi: 10.1007/978-3-031-19806-9_22. [14] SONG Zeyin, ZHAO Yifan, SHI Yujun, et al. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning[C]. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, Canada, 2023: 24183–24192. doi: 10.1109/CVPR52729.2023.02316. [15] WANG Li, YANG Xinyao, TAN Haoyue, et al. Few-shot class-incremental SAR target recognition based on hierarchical embedding and incremental evolutionary network[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5204111. doi: 10.1109/TGRS.2023.3248040. [16] ZHAO Yan, ZHAO Lingjun, DING Ding, et al. Few-shot class-incremental SAR target recognition via cosine prototype learning[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5212718. doi: 10.1109/TGRS.2023.3298016. [17] NOVAK L M, OWIRKA G J, BROWER W S, et al. The automatic target-recognition system in SAIP[J]. The Lincoln Laboratory Journal, 1997, 10(2): 187–202. [18] 王智睿, 康玉卓, 曾璇, 等. SAR-AIRcraft-1.0: 高分辨率SAR飛機(jī)檢測(cè)識(shí)別數(shù)據(jù)集[J]. 雷達(dá)學(xué)報(bào), 2023, 12(4): 906–922. doi: 10.12000/JR23043.WANG Zhirui, KANG Yuzhuo, ZENG Xuan, et al. SAR-AIRcraft-1.0: High-resolution SAR aircraft detection and recognition dataset[J]. Journal of Radars, 2023, 12(4): 906–922. doi: 10.12000/JR23043. -