Insufficiently labeled samples and low-generalization performance have become significant natural language processing problems, drawing significant concern for few-shot text classification (FSTC). Advances in prompt learning have significantly improved the performance of FSTC. However, prompt learning methods typically require the pre-trained language model and tokens of the vocabulary list for model training, while different language models have different token coding structures, making it impractical to build effective Chinese prompt learning methods from previous approaches related to English. In addition, a majority of current prompt learning methods do not make use of existing unlabeled data, thus often leading to unsatisfactory performance in real-world applications. To address the above limitations, we propose a novel Chinese FSTC method called CIPLUD that combines an improved prompt learning method and existing unlabeled data, which are used for the classification of a small amount of Chinese text data. We used the Chinese pre-trained language model to build two modules: the Multiple Masks Optimization-based Prompt Learning (MMOPL) module and the One-Class Support Vector Machine-based Unlabeled Data Leveraging (OCSVM-UDL) module. The former generates prompt prefixes with multiple masks and constructs suitable prompt templates for Chinese labels. It optimizes the random token combination problem during label prediction with joint probability and length constraints. The latter, by establishing an OCSVM model in the trained text vector space, selects reasonable pseudo-label data for each category from a large amount of unlabeled data. After selecting the pseudo-label data, we mixed them with the previous few-shot annotated data to obtain brand new training data and then repeated the steps of the two modules as an iterative semi-supervised optimization process. The experimental results on the four Chinese FSTC benchmark datasets demonstrate that our proposed solution outperformed other prompt learning methods with an average accuracy improvement of 2.3%.
2 October 2025 in IEEE Transactions on Aerospace and Electronic Systems
Millimeter-wave Synthetic Aperture Radar (SAR) imaging is widely used in security screening, industrial nondestructive testing, and autonomous navigation due to its high resolution and penetration capabilities. However, high-fidelity reconstruction typically requires dense spatial sampling, resulting in long scanning times and significant data acquisition overhead. Although compressed sensing has been explored for sparse reconstruction, its reliance on random sampling is incompatible with the deterministic scanning trajectories of practical SAR systems and thus fails to reduce real acquisition time. To address this, we propose mmUSAR, an unsupervised SAR imaging framework incorporating two hardware-compliant sparse scanning strategies: Expanded Vertical-interval Sampling (EVS) and Reduced Height Sampling (RHS), which reduce the number of vertical scanning positions. Reconstruction is performed by an untrained complexvalued neural network that embeds physical priors via an observation approximation operator and enforces structured sparsity regularization. The framework requires no external training data and adapts to each scene through test-time optimization. Experiments on simulated and real-world datasets show that mmUSAR achieves high-fidelity image reconstruction while reducing the scanning trajectory, thereby decreasing acquisition time by up to 75%. Compared to conventional and other untrained neural network methods, mmUSAR offers superior image quality and competitive runtime, providing a practical solution for efficient and resource constrained SAR imaging.
Direction-of-arrival (DOA) estimation from a single snapshot becomesconsiderably harder when sensors in the array fail, yet reliable angularmeasurement under these conditions is critical for automotive radar, sonar,and acoustic surveillance. Random antenna outagesdestroy the covariance structure relied upon by subspace methods and introducesample-dependent aperture sparsity that conventional array processing cannothandle. This paper proposes MC-CoDA, a failure-aware deep learningarchitecture that propagates sensor health information throughout the entireestimation pipeline. The model constructs sensor-aligned tokens from complexarray observations, injects multi-scale local features through a cascadeconvolutional branch, regulates cross-sensor interaction via mask-guidedattention with a learnable geometry--health bias, and performs failure-awareglobal aggregation through mask-weighted pooling. Experiments on a uniformlinear array show that MC-CoDA delivers sub-degree accuracy acrossmoderate-to-high SNR even under heavy random sensor outages, sustains reliabletwo-source resolution at small angular separations, and exhibits the slowestdegradation among all evaluated methods as the failure ratio rises beyond thetraining cap.Ablation results identify the mask-guided attention and mask-weighted poolingas the primary sources of robustness, while the cascade local branchcontributes a complementary gain in their presence. Evaluation on measuredradar data confirms that the learned failure-awareness transfers to practicalarrays subject to unmodeled clutter and mutual coupling.