.png){kind=icon}
.png){kind=icon}
Return
Multitask Joint Optimization Based SAR Target Recognition Model From Echo Data
.png){kind=icon}
OA
AI
DOI：10.1109/JSTARS.2026.3673000.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
To address the limitations of conventional synthetic aperture radar (SAR) target recognition, particularly the computational redundancy and error accumulation inherent in the imaging-first paradigm, the paper proposes echo-to-class network (E2CNet), an end-to-end recognition model operating directly on echo data via multitask joint optimization. By bypassing intermediate imaging procedures, the proposed E2CNet directly maps echo data to target categories. Central to the E2CNet architecture is a deep complex-valued neural network designed to preserve and leverage scattering characteristics within the echo data. A bidirectional long short-term memory (Bi-LSTM) network is subsequently integrated to capture spatiotemporal dependencies in the feature sequences. Furthermore, a scattering point adaptive extraction strategy, guided by SAR image priors, is incorporated into a multitask learning framework to enhance both recognition accuracy and scattering center localization precision. Experiments on multicategory echo datasets derived from the MSTAR and FUSAR-Ship benchmarks demonstrate that the E2CNet achieves competitive recognition performance, attaining accuracies of 86.45% and 91.60%, respectively. Moreover, validations on real-world data from the Fucheng-1 satellite further confirm the recognition capability of the model and its advantage in reducing end-to-end inference time. The results highlight the potential of E2CNet for efficient and accurate SAR target recognition in practical applications.
Keywords:
End-to-end
multitask joint optimization
synthetic aperture radar (SAR)
target recognition
Journal
IF:
5.3
Papers:
1.3K
Citations:
3.0W